chore: v0.1.4 — 版本号收拢（6 处）

internal/version/version.go（ldflags 示例 + 默认值）、scripts/release.sh（注释 + 默认值）、
scripts/build_installers.sh、packaging/rpm/webui4frpc.spec（Version + changelog，含本次
负载均衡四条）、README.md 与 packaging/windows/README.md 的安装示例。

漏一处就会让某个安装包的示例文件名指向不存在的产物，所以按 v0.1.3 的 5 处清单逐个核对，
并补上 084ef62 才纳入的 packaging/windows/README.md，共 6 个文件。

README.md

@@ -32,6 +32,14 @@
   - 主负载信号：当前节点拥有的转发数（权重 ×100）
   - 次级信号：**NIC 网络接口利用率**（从 `/proc/net/dev` 增量 + `/sys/class/net/*/speed` 计算，排除了 lo/docker/veth 等虚拟接口）
   - 同转发数节点间由内存占用 + 网络饱和度打破平局
+- **负载均衡（自导式）**：每个节点对**自己的**转发跑同一条规则 —— 持有数超过
+  `ceil(总转发数 / 存活节点数)` 且至少 2 条时，自行撤销一条、把任务丢回环上，
+  再由环按「最闲者优先」重新认领
+  - **没有中心调度**：一节点的负载只有它自己的拓扑条目消失时才会下降，而条目是
+    本地簿记，没有对端能代它写，所以只能由持有者自己动手
+  - 规则自限：撤到平摊值即停（否则会把转发卸光），且低于下限永不撤、不会卸空节点
+  - 每节点默认 10 分钟冷却，限制抖动频率；默认 **dry-run**（只报告不释放）
+  - 集群页每节点一个开关：关闭 / 观察模式 / 启用，可持久化
 - **per-forward 独立进程**：每条转发拥有独立的 frpc 配置文件和 frpc 进程（worker 键 = `local~remote~port`），单条转发故障不影响兄弟，消除旧模型的多节点 `proxy already exists` 冲突
 - **容错自愈**：leader / 非 leader 节点宕机 → 自动检测 → 环跳过死亡节点、任务重挂；leader 宕机 → 上邻居晋升为新 leader
 - **宕机自动重联**：节点宕机重启后通过缓存的集群拓扑（对端地址 + 准入密钥）自动重联，显式脱离集群的节点不自动重联
@@ -79,7 +87,7 @@ cd web && npm run build && rm -rf internal/httpapi/dist && cp -r web/dist intern
 ### Linux (Debian / Ubuntu)
 
 ```bash
-sudo dpkg -i webui4frpc-0.1.1-linux-amd64.deb   # 或 linux-arm64.deb
+sudo dpkg -i webui4frpc-0.1.4-linux-amd64.deb   # 或 linux-arm64.deb
 sudo nano /etc/webui4frpc/webui4frpc.env        # 改掉默认密码
 sudo systemctl restart webui4frpc
 ```
@@ -87,7 +95,7 @@ sudo systemctl restart webui4frpc
 ### Linux (RHEL / Fedora / openEuler / 默认 / 龙蜥)
 
 ```bash
-sudo rpm -i webui4frpc-0.1.1-linux-x86_64.rpm
+sudo rpm -i webui4frpc-0.1.4-linux-x86_64.rpm
 sudo vi /etc/webui4frpc/webui4frpc.env
 sudo systemctl enable --now webui4frpc
 ```
@@ -97,7 +105,7 @@ sudo systemctl enable --now webui4frpc
 
 ### Windows
 
-双击 `webui4frpc-0.1.1-windows-amd64-setup.exe`（ARM 设备用 `windows-arm64-setup.exe`）。
+双击 `webui4frpc-0.1.4-windows-amd64-setup.exe`（ARM 设备用 `windows-arm64-setup.exe`）。
 安装向导会询问监听地址与管理员凭据，完成后：
 
 - 程序装到 `C:\Program Files\webui4frpc`
@@ -107,8 +115,8 @@ sudo systemctl enable --now webui4frpc
 ### macOS
 
 ```bash
-sudo installer -pkg webui4frpc-0.1.1-darwin-arm64.pkg -target /   # Apple Silicon
-sudo installer -pkg webui4frpc-0.1.1-darwin-amd64.pkg -target /   # Intel
+sudo installer -pkg webui4frpc-0.1.4-darwin-arm64.pkg -target /   # Apple Silicon
+sudo installer -pkg webui4frpc-0.1.4-darwin-amd64.pkg -target /   # Intel
 ```
 
 或直接双击 `.pkg` 走图形安装器。安装后 launchd 开机自启，配置在
@@ -126,7 +134,7 @@ sudo launchctl kickstart -k system/com.jianfeeeee.webui4frpc
 不想装服务只想跑一下，用对应的 `.tar.gz` / `.zip`，解压即用：
 
 ```bash
-tar xzf webui4frpc-0.1.1-linux-amd64.tar.gz
+tar xzf webui4frpc-0.1.4-linux-amd64.tar.gz
 ./webui4frpc -addr 127.0.0.1:7500 -user admin -password 你的密码 -workdir ./data
 ```
 
@@ -160,6 +168,19 @@ sha256sum -c SHA256SUMS --ignore-missing
 
 环境变量 `W4F_HOST` 可覆盖节点对外可达的主机名（Docker DNS / 通配监听场景用）。
 
+### 负载均衡参数
+
+每节点独立生效；集群页的开关**优先于**这里的默认值（在页面上选过就以页面为准，重启不丢）。
+
+| 环境变量 | 默认 | 说明 |
+| --- | --- | --- |
+| `W4F_BALANCE_MODE` | `dry-run` | `off` 不参与 / `dry-run` 只报告 / `on` 真正释放 |
+| `W4F_BALANCE_COOLDOWN` | `600` | 同一节点两次释放之间的最小秒数 |
+| `W4F_BALANCE_INTERVAL` | `10` | 检查间隔（秒） |
+
+**释放会停掉该转发的 frpc worker，其上的在途连接会断一次**，由新节点认领后恢复；
+因此默认是 `dry-run`。集群页的开关走 write 级权限。
+
 ### 启动行为
 
 节点启动时的集群引导优先级：
@@ -191,10 +212,10 @@ webui4frpc（单二进制，无 frp 依赖，前端 go:embed）
 │   ├── render               渲染 frpc JSON 配置（per-forward 单 proxy + 高级参数 + LB/健康检查）
 │   ├── process              worker 进程管理（spawn/stop/restart/自愈/日志轮转，worker 键 = local~remote~port）
 │   ├── install              一键下载官方 frpc（GitHub Releases，含解压安全防护）
-│   ├── cluster              令牌环协议（ring engine / leader 选举 / 容错 / 宕机重联 / 二进制分发 / NIC 负载采样）
+│   ├── cluster              令牌环协议（ring engine / leader 选举 / 容错 / 宕机重联 / 二进制分发 / NIC 负载采样 / 自导式负载均衡）
 │   └── httpapi              REST API + 静态资源（auth/session / canvas / forwards / cluster / users / audit 导出）
 └── web/                     Vue 3 SPA（Element Plus + VueFlow 画布 + Pinia）
-    └── src/views/           状态总览 / 连接配置（画布）/ 集群 / 设置 / 账号与密钥
+    └── src/views/           状态总览 / 连接配置（画布）/ 集群（含负载均衡开关）/ 设置 / 账号与密钥
 ```
 
 ## API
@@ -220,6 +241,8 @@ webui4frpc（单二进制，无 frp 依赖，前端 go:embed）
 | | GET | `/binary/status` | read | frpc 路径 |
 | | POST | `/binary/install` | write | 安装 frpc |
 | **集群** | GET | `/cluster/ring` | read | 令牌环快照 |
+| | GET | `/cluster/balance` | read | 本节点负载均衡视图（只读，不变更状态） |
+| | POST | `/cluster/balance/mode` | write | 切换本节点负载均衡模式（持久化） |
 | | POST | `/cluster/create` | write | 创建独立集群 |
 | | POST | `/cluster/join-ring` | write | 加入集群（本节点发起） |
 | | POST | `/cluster/join` | write | 接收新节点加入 |
@@ -254,8 +277,45 @@ webui4frpc（单二进制，无 frp 依赖，前端 go:embed）
 - **任务分发**：在任意节点画布上配置 `localOnly=false` 的转发并保存 → 画布差异判断生成命令 → 令牌携带 → 负载最低节点摘取 → 拉起 frpc worker
 - **容错**：节点宕机 → 自动检测 → 环自愈（跳过死亡节点 / leader failover）；重启 → 自动重联
 - **退出集群**：在集群页「退出集群」→ 本节点脱离为 standalone，不自动重联
+- **负载均衡**：见下节
 
-集群页展示：环拓扑（self/leader 标记）、令牌轮次、节点负载、待办任务、活跃转发拓扑（含 owner 归属）、增量事件日志。
+集群页展示：环拓扑（self/leader 标记）、令牌轮次、节点负载、待办任务、活跃转发拓扑（含 owner 归属）、负载均衡（本机份额与释放计划）、增量事件日志。
+
+### 负载均衡（自导式释放）
+
+**为什么不是 leader 统一调度**：一节点的负载信号是它持有的转发数，而这个数只在
+**它自己的拓扑条目消失**时才下降 —— 条目是本地簿记。所以 leader 从外面只能*看见*
+失衡，无法*消除*它；要消除就得伸手改另一个节点的簿记，而这正是这个环到处都在避免
+的跨节点写（撤销、重启同样都是定向往 owner）。
+
+于是规则由每个节点对自己的转发本地执行：
+
+```
+当  持有转发数 >= 2  且  持有转发数 > ceil(总转发数 / 存活节点数)
+则  撤销其中一条 → 删除条目 → 以启用状态重新发布到环 → 由最闲节点认领
+```
+
+- **自限**：撤到平摊值即停。3 条分 3 台时，持有 2 条的那台卸一条后变成 1 条，
+  `1 > 1` 不成立 → 停手。没有这个上界，owner 会把转发全部卸光。
+- **不会卸空**：低于下限 2 条永不撤销。
+- **不碰已停用的转发**：停用条目的 entry 是用户「停用」决定的载体，释放它等于把
+  一条**已停**的转发当活干交给别人。
+- **释放 ≠ 停止**：用户停止保留条目并标记链路为停用；均衡释放**删除条目**且链路保持
+  启用。两者不能混用 —— 复用停止的话，每个节点的重复认领守卫都会拒绝这条已是「有主」
+  的转发，它会被永久搁死在刚放弃它的那台机器上。
+
+三种模式（集群页每节点一个开关，也可用 `W4F_BALANCE_MODE` 设默认值）：
+
+| 模式 | 行为 |
+| --- | --- |
+| `off` | 不参与均衡 |
+| `dry-run` | 只报告「将释放哪条」，不动任何东西（默认） |
+| `on` | 真正释放 |
+
+只读接口 `GET /cluster/balance` 返回本节点视角：`selfForwards` / `share` /
+`totalForwards` / `aliveNodes` / `selfOverShare`、`wouldRelease`、`releasing`、
+`blockedBy`（为何暂不动的机器原因）、`mode` / `modeSource` / `canControl`。
+该接口**不改变任何状态** —— 反复轮询不会启动、延迟或取消一次释放。
 
 ## 测试
 
@@ -263,6 +323,9 @@ webui4frpc（单二进制，无 frp 依赖，前端 go:embed）
 # 后端
 go test ./...
 
+# 单包（负载均衡规则与状态机）
+go test ./internal/cluster/ -run TestShare -v
+
 # 前端类型检查
-cd web && npm run type-check
+cd web && npm run build
 ```

cmd/webui4frpc/main.go

@@ -15,6 +15,7 @@ import (
 	"os"
 	"os/signal"
 	"path/filepath"
+	"strconv"
 	"strings"
 	"syscall"
 	"time"
@@ -175,13 +176,41 @@ func main() {
 					return err
 				}
 				// A re-claim after a forward-centric stop leaves the link flagged
-				// disabled (by RevokeFn); clear it so renderRemote renders the
-				// proxy back in. No-op for a fresh claim.
-				_ = st.SetLinkDisabled(tk.Local.Name, tk.Remote.Name, tk.Link.RemotePort, false)
+				// disabled (by RevokeFn). Do NOT clear that flag here: the ring
+				// re-claims automatically (startup reconcile re-spawns any owned
+				// forward whose worker is missing), so clearing it on claim made
+				// "stopped" un-durable — every restart resurrected a forward the user
+				// had explicitly stopped, and it then failed forever against a local
+				// service that was intentionally not running.
+				//
+				// The flag is now owned by the two user-facing actions:
+				//   startForward → SetLinkDisabled(false) before submitting the task
+				//   stopForward  → SetLinkDisabled(true)  before revoking it
+				// Claim only READS it to decide whether to bring the worker up. A fresh
+				// claim of a link with no row still starts, because the lookup miss
+				// below is treated as "not disabled".
+				disabled := false
+				if existing, found, err := st.LinkByTriple(tk.Link.Local, tk.Link.Remote, tk.Link.RemotePort); err != nil {
+					return err
+				} else if found {
+					disabled = existing.Disabled
+				}
 				// Start (or restart) the per-forward worker for exactly this link.
 				// Each forward has its own frpc process (keyed by the forward
 				// triple); restarting only this key leaves sibling forwards'
 				// processes untouched.
+				if disabled {
+					// Mark the entry stopped rather than deleting it. The entry is the
+					// carrier that keeps the flag travelling around the ring, and
+					// UpdateTopologyDisabled also clears Active — which is what makes
+					// this entry inert for OfflineReassign() and the startup
+					// reconcile, so it is not resurrected later.
+					if ring != nil {
+						ring.UpdateTopologyDisabled(tk.Local.Name, tk.Remote.Name, tk.Link.RemotePort, true)
+					}
+					log.Printf("ring[%s] claim %s skipped: %s→%s:%d is disabled", selfID, tk.ID, tk.Local.Name, tk.Remote.Name, tk.Link.RemotePort)
+					return nil
+				}
 				if tk.Remote.Enabled {
 					key := process.WorkerKey(tk.Local.Name, tk.Remote.Name, tk.Link.RemotePort)
 					if _, has := pm.Status(key); has {
@@ -201,13 +230,118 @@ func main() {
 			// stop/start: it wiped sibling forwards sharing the local or remote.
 			RevokeFn: func(ctx context.Context, tk *cluster.Task) error {
 				_ = st.SetLinkDisabled(tk.Local.Name, tk.Remote.Name, tk.Link.RemotePort, true)
+				// The links table is node-local: only the node that received the
+				// stop request has the flag persisted, while the node that OWNS the
+				// forward is the one holding the worker (and is often a different
+				// machine). Trusting only the local row meant the owner's copy could
+				// still read disabled=false. The task now carries the flag, so use it
+				// as the authority and write it here.
+				ln, found, err := st.LinkByTriple(tk.Local.Name, tk.Remote.Name, tk.Link.RemotePort)
+				switch {
+				case err != nil:
+					return err
+				case found:
+					ln.Disabled = true
+					// Persist via the same rewrite path the canvas uses, so the flag
+					// survives a restart on THIS node too.
+					links, err := st.ListLinks()
+					if err != nil {
+						return err
+					}
+					for i := range links {
+						if links[i].Local == ln.Local && links[i].Remote == ln.Remote && links[i].RemotePort == ln.RemotePort {
+							links[i].Disabled = true
+						}
+					}
+					if err := st.ReplaceLinks(links); err != nil {
+						return err
+					}
+				case tk.Link.Disabled:
+					// No local row yet (this node never claimed the forward) but the
+					// task asserts the stop — record it so a later claim on this node
+					// cannot resurrect the forward.
+					loc, rem := tk.Local, tk.Remote
+					links, err := st.ListLinks()
+					if err != nil {
+						return err
+					}
+					links = append(links, store.Link{
+						Local: loc.Name, Remote: rem.Name, RemotePort: tk.Link.RemotePort,
+						OffsetX: tk.Link.OffsetX, OffsetY: tk.Link.OffsetY,
+						Group: tk.Link.Group, Disabled: true,
+					})
+					if err := st.ReplaceLinks(links); err != nil {
+						return err
+					}
+				}
 				key := process.WorkerKey(tk.Local.Name, tk.Remote.Name, tk.Link.RemotePort)
 				if _, running := pm.Status(key); running {
 					_ = pm.Stop(key)
 				}
+				// Mark the entry stopped instead of deleting it: the entry carries
+				// the flag around the ring, and clearing Active keeps it inert for
+				// the rebalancing paths (OfflineReassign skips inactive entries, and
+				// the startup reconcile skips disabled forwards), so the worker is
+				// not re-spawned on the next restart. Deleting it here is what used
+				// to leave the stop with no carrier at all.
+				if ring != nil {
+					ring.UpdateTopologyDisabled(tk.Local.Name, tk.Remote.Name, tk.Link.RemotePort, true)
+				}
 				log.Printf("ring[%s] revoked task %s: %s→%s:%d", selfID, tk.ID, tk.Local.Name, tk.Remote.Name, tk.Link.RemotePort)
 				return nil
 			},
+			// ReleaseFn stops this forward's worker for a REBALANCE, leaving the
+			// link ENABLED. This must not reuse RevokeFn: that marks the link
+			// stopped, and the released forward is about to run on another node —
+			// a "stopped" flag would make the claim path treat the re-published
+			// task as a disabled forward and refuse to start a worker for it,
+			// leaving the forward dead everywhere.
+			ReleaseFn: func(ctx context.Context, tk *cluster.Task) error {
+				key := process.WorkerKey(tk.Local.Name, tk.Remote.Name, tk.Link.RemotePort)
+				if _, running := pm.Status(key); running {
+					_ = pm.Stop(key)
+				}
+				log.Printf("ring[%s] released worker %s (rebalancing): %s→%s:%d", selfID, key,
+					tk.Local.Name, tk.Remote.Name, tk.Link.RemotePort)
+				return nil
+			},
+			// RestartFn brings an already-claimed forward back after a stop. It
+			// must NOT run the claim bookkeeping (ClaimFn), which appends links and
+			// writes topology state: the forward is already established, and a stop
+			// deliberately keeps its topology entry, so all that is needed is to
+			// clear the flag and respawn the worker. The engine has already
+			// re-marked the topology entry enabled before calling this.
+			RestartFn: func(ctx context.Context, tk *cluster.Task) error {
+				// Clear the stop on this node's own store, using the same path the
+				// start handler uses, so the local copy cannot disagree with the
+				// ring and block a later restart.
+				if ln, found, err := st.LinkByTriple(tk.Link.Local, tk.Link.Remote, tk.Link.RemotePort); err != nil {
+					return err
+				} else if found && ln.Disabled {
+					links, err := st.ListLinks()
+					if err != nil {
+						return err
+					}
+					for i := range links {
+						if links[i].Local == ln.Local && links[i].Remote == ln.Remote && links[i].RemotePort == ln.RemotePort {
+							links[i].Disabled = false
+						}
+					}
+					if err := st.ReplaceLinks(links); err != nil {
+						return err
+					}
+				}
+				if tk.Remote.Enabled {
+					key := process.WorkerKey(tk.Local.Name, tk.Remote.Name, tk.Link.RemotePort)
+					if _, has := pm.Status(key); has {
+						_ = pm.Restart(key)
+					} else {
+						_ = pm.Start(key)
+					}
+				}
+				log.Printf("ring[%s] restarted %s: %s→%s:%d", selfID, tk.ID, tk.Local.Name, tk.Remote.Name, tk.Link.RemotePort)
+				return nil
+			},
 		},
 		ringTransport.SendTo(func(nodeID string) string {
 			if ring == nil {
@@ -232,10 +366,24 @@ func main() {
 	// left does NOT auto-rejoin.
 	ring.SetPeerPersist(func(peersJSON string) error { return st.SetClusterPeers(peersJSON) })
 
-	// Re-apply local store group labels onto the ring topology after each
-	// state adoption. Without this, group changes made via HTTP handlers
+	// Re-apply local store state onto the ring topology after each state
+	// adoption, and learn the ring's disabled decisions back into the store.
+	//
+	// Without the group half, group changes made via HTTP handlers
 	// (POST /forwards/assign) are overwritten by the next e.state = tk.State
-	// and never propagate to other nodes.
+	// and never propagate.
+	//
+	// The disabled half makes the RING authoritative for stop/start. That is the
+	// only source all members can agree on: the links table is a per-node copy, so
+	// treating it as authoritative let two nodes disagree indefinitely (observed
+	// live: one node reading disabled=1 while the owner still read 0 and kept its
+	// worker running). This is a write-through, not a listener — it runs on every
+	// adoption, so any peer's decision lands here within a token cycle.
+	//
+	// Ordering matters in one direction only: the engine has already re-applied
+	// this node's not-yet-confirmed local decisions onto the adopted state (see
+	// reconcileLocalDisabled), so reading the ring here can never clobber an
+	// action the user just took on this node.
 	ring.SetTopologySync(func() {
 		links, err := st.ListLinks()
 		if err != nil {
@@ -244,8 +392,23 @@ func main() {
 		for _, ln := range links {
 			ring.UpdateTopologyGroup(ln.Local, ln.Remote, ln.RemotePort, ln.Group)
 		}
+		for _, te := range ring.State().TopologyList() {
+			disabled, known := ring.TopologyDisabled(te.Local.Name, te.Remote.Name, te.Link.RemotePort)
+			if !known {
+				continue
+			}
+			if ln, found, err := st.LinkByTriple(te.Local.Name, te.Remote.Name, te.Link.RemotePort); err == nil && found && ln.Disabled == disabled {
+				continue // store already agrees with the ring
+			}
+			_ = st.ReconcileLinkDisabled(te.Local.Name, te.Remote.Name, te.Link.RemotePort, disabled)
+		}
 	})
 
+	// Load balancer is created here (before the Handler) so the HTTP layer can
+	// expose its read-only report. Mode/parameters are applied further down, once
+	// the W4F_BALANCE_* environment has been parsed.
+	balancer := cluster.NewBalancer(ring, 600)
+
 	h := &httpapi.Handler{
 		Store:    st,
 		Process:  pm,
@@ -254,6 +417,9 @@ func main() {
 		Sessions: httpapi.NewSessionStore(),
 		Cluster:  reg,
 		Ring:     ring,
+		// Balancer is created above, before the environment that configures it is
+		// parsed, because the HTTP layer needs the pointer.
+		Balancer: balancer,
 		// SelfAddr must be reachable by peers. 0.0.0.0/empty on a real host
 		// would break peer reconnect, so fall back to the hostname (container
 		// name inside docker is a routable addr on the compose network).
@@ -411,6 +577,65 @@ func main() {
 	go ring.WatchTokenLoss(clusterCtx)
 	go ring.WatchLeader(clusterCtx)
 
+	// Load balancing — SELF-DIRECTED, and dry-run by default.
+	//
+	// Every node runs this check against its OWN forwards. There is no
+	// coordinator: a node's load only actually falls when its own topology entry
+	// goes away, so only the owner can produce the state change that makes the
+	// ring hand a forward to somebody else. The node gives a forward up and
+	// re-publishes it as an ordinary unclaimed task; the ring's normal
+	// lowest-load claim then places it. (The earlier design had the leader pick a
+	// victim and a destination — but it can only see the imbalance, not resolve
+	// it, so it had to reach into another node's bookkeeping to make its decision
+	// true.)
+	//
+	// A release stops a worker and drops a live connection, so acting on the ring
+	// is opt-in:
+	//
+	//   W4F_BALANCE_MODE=dry-run   report the decision, release nothing (default)
+	//   W4F_BALANCE_MODE=on        actually release
+	//   W4F_BALANCE_MODE=off       do nothing at all
+	//   W4F_BALANCE_COOLDOWN=600   min seconds between two releases by one node
+	//   W4F_BALANCE_INTERVAL=10    check every N seconds
+	balMode := strings.ToLower(strings.TrimSpace(os.Getenv("W4F_BALANCE_MODE")))
+	if balMode == "" {
+		balMode = cluster.BalanceDryRun
+	}
+	balInterval := time.Duration(envFloat("W4F_BALANCE_INTERVAL", 10)) * time.Second
+	if balInterval < time.Second {
+		balInterval = 10 * time.Second
+	}
+	balCooldown := int64(envFloat("W4F_BALANCE_COOLDOWN", 600))
+	balancer.SetCooldown(balCooldown)
+	balancer.SetMode(balMode)
+	// A mode stored from the cluster page wins over the environment, so flipping
+	// the switch is not silently reverted by the next restart. SetMode is applied
+	// first so a stored value only has to be a no-op when the two agree.
+	balancer.SetModePersistence(
+		func() (string, bool) {
+			m, ok, err := st.BalanceMode()
+			if err != nil {
+				log.Printf("ring[%s] read stored balance mode: %v", selfID, err)
+				return "", false
+			}
+			return m, ok
+		},
+		func(m string) error {
+			s, err := st.Settings()
+			if err != nil {
+				return err
+			}
+			s.BalanceMode = m
+			return st.UpdateSettings(s)
+		},
+	)
+	// Report the EFFECTIVE mode, not balMode: a stored value may have overridden
+	// the environment, and a startup line naming the wrong mode is exactly how an
+	// operator ends up believing releasing is off while it is on.
+	log.Printf("ring[%s] load balancing %s (source=%s cooldown=%ds interval=%s, self-directed)",
+		selfID, strings.ToUpper(balancer.Mode()), balancer.ModeSource(), balCooldown, balInterval)
+	go balancer.Watch(clusterCtx, balInterval)
+
 	// Startup reconcile: every node (leader or not) checks its own topology
 	// entries and re-claims any whose per-forward frpc worker process is
 	// absent — the topology was recovered via AdoptState but the worker
@@ -419,17 +644,32 @@ func main() {
 	// LocalOnly forwards are never in the topology and are handled by the
 	// auto-start loop above.
 	go func() {
-		for {
-			if ring.IsLeader() {
-				break
-			}
+		// No leader gate here. The reconcile body is per-NODE work: it re-claims
+		// the forwards THIS node owns, whose worker processes died with the
+		// restart. Every member owns forwards (a 3-node ring splits them), so
+		// gating on leadership meant non-leaders waited forever and never
+		// rebuilt their own workers — a node could come back up "in the cluster"
+		// while owning a forward with no process behind it. Observed live: after a
+		// rolling restart a non-leader rejoined, adopted state listing itself as
+		// the owner of a forward, and never started it, because it sat waiting to
+		// become leader.
+		//
+		// What actually needs settling is the ring, not the leadership: wait for
+		// the topology to show us (with a bounded fallback so a lone node that
+		// rejoined a ring whose leader is unreachable still makes progress).
+		deadline := time.After(30 * time.Second)
+		for !ring.HasSelfInTopology() {
 			select {
-			case <-time.After(2 * time.Second):
+			case <-deadline:
+				log.Printf("ring[%s] startup reconcile: self not in topology after 30s, proceeding anyway", ring.ID)
+			case <-time.After(500 * time.Millisecond):
+				continue
 			case <-clusterCtx.Done():
 				return
 			}
+			break
 		}
-		// Let the ring stabilize after leader election / rejoin.
+		// Let the ring stabilize after election / rejoin before acting on it.
 		time.Sleep(3 * time.Second)
 		snap := ring.Snapshot()
 		reclaimed := 0
@@ -437,6 +677,16 @@ func main() {
 			if t.OwnerID != selfID {
 				continue
 			}
+			// A forward the user stopped must stay stopped: skip it here so a
+			// restart does not re-spawn its worker. The ClaimFn enforces the
+			// same rule (belt and braces — this also avoids a pointless
+			// claim→skip round trip per disabled forward on every boot).
+			if ln, found, err := st.LinkByTriple(t.Local.Name, t.Remote.Name, t.Link.RemotePort); err != nil {
+				log.Printf("ring[%s] reconcile: lookup %s→%s:%d: %v", ring.ID, t.Local.Name, t.Remote.Name, t.Link.RemotePort, err)
+				continue
+			} else if found && ln.Disabled {
+				continue
+			}
 			key := process.WorkerKey(t.Local.Name, t.Remote.Name, t.Link.RemotePort)
 			if _, has := pm.Status(key); has {
 				continue // worker already running
@@ -568,6 +818,22 @@ func retryJoinCluster(ctx context.Context, ring *cluster.Engine, targetAddr, joi
 	}
 }
 
+// envFloat reads a numeric environment variable, returning def when unset or
+// unparseable. A malformed value must not silently change behaviour, so it falls
+// back to the documented default rather than 0.
+func envFloat(key string, def float64) float64 {
+	v := strings.TrimSpace(os.Getenv(key))
+	if v == "" {
+		return def
+	}
+	f, err := strconv.ParseFloat(v, 64)
+	if err != nil {
+		log.Printf("env %s=%q is not a number; using %v", key, v, def)
+		return def
+	}
+	return f
+}
+
 // retryRejoinCached is the crash/restart recovery path: parse the persisted
 // peer list (JSON of [{addr,key},...]) and try to rejoin via each peer in
 // order. The first peer that accepts wins. Retry the whole list every 10s
@@ -600,6 +866,20 @@ func retryRejoinCached(ctx context.Context, ring *cluster.Engine, peersJSON stri
 			cancel()
 			if err == nil {
 				log.Printf("ring[%s] rejoined cluster via cached peer %s", ring.ID, p.Addr)
+				// A rejoined node must be able to drive the ring itself. The
+				// normal case is that the existing leader keeps circulating the
+				// token, but if WE are the leader (typical after a rolling
+				// restart where the leader came back last, or the leader died
+				// while we were down) nobody else will: StartRing is a no-op on
+				// non-leaders, and the ring sits still with every task stranded
+				// in pending. Observed live: the leader restarted, rejoined fine,
+				// and then never forwarded a token — cycle frozen, topology
+				// empty, pending=3.
+				//
+				// StartRing is safe to call unconditionally: it returns
+				// immediately unless this node leads, and it throttles itself
+				// against RoundDelay so a redundant call cannot double-send.
+				go ring.StartRing(ctx)
 				return
 			}
 			log.Printf("ring[%s] rejoin %s failed: %v", ring.ID, p.Addr, err)

cmd/webui4frpc/rejoin_ring_test.go

@@ -0,0 +1,145 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"webui4frpc/internal/cluster"
+	"webui4frpc/internal/store"
+)
+
+// fakeJoinPeer stands in for a healthy cluster member: it answers the rejoin
+// POST and records it, so the rejoin path can be exercised end to end.
+type fakeJoinPeer struct {
+	mu      sync.Mutex
+	joined  int
+	stateFn func() cluster.State
+}
+
+func (p *fakeJoinPeer) ServeHTTP(w http.ResponseWriter, r *http.Request) {
+	p.mu.Lock()
+	p.joined++
+	p.mu.Unlock()
+	w.Header().Set("Content-Type", "application/json")
+	_ = json.NewEncoder(w).Encode(map[string]any{"state": p.stateFn()})
+}
+
+func (p *fakeJoinPeer) joins() int {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return p.joined
+}
+
+// TestRejoinThenStartRingCoversTheRejoinPath is the regression test for the
+// frozen ring, and it deliberately drives retryRejoinCached itself rather than
+// calling StartRing directly — a test that calls the fix's own helper proves
+// nothing about whether the fix is wired up (verified: such a test stayed green
+// with the fix reverted).
+//
+// The failure, observed live on the three-node cluster: the leader restarted,
+// rejoined successfully, logged "startup reconcile: reclaimed=0 submitted=3",
+// and then never forwarded a token. cycle froze, topology stayed empty and the
+// three forwards sat in pending forever, because rejoin returned immediately
+// and only a leader may call StartRing.
+func TestRejoinThenStartRingCoversTheRejoinPath(t *testing.T) {
+	selfID := "192.168.2.30:7500"
+
+	// The peer hands back a ring in which SELF is the leader — the situation a
+	// rolling restart produces when the leader comes back last.
+	var peerState cluster.State
+	peer := &fakeJoinPeer{stateFn: func() cluster.State { return peerState }}
+
+	// Bind on loopback so the engine's hardcoded peer address is reachable.
+	// We therefore advertise 127.0.0.1 in the ring while keeping the same IDs.
+	ln := httptest.NewServer(peer)
+	defer ln.Close()
+
+	peerState = cluster.State{
+		LeaderID: selfID,
+		Nodes: []cluster.Node{
+			{ID: selfID, Addr: selfID, Alive: true, IsLeader: true},
+			{ID: "peer:7500", Addr: "peer:7500", Alive: true},
+		},
+		Topology: map[string]*cluster.TopoEntry{},
+		PendingTasks: map[string]*cluster.Task{
+			"t1": {ID: "t1", Local: store.Local{Name: "web"},
+				Remote: store.Remote{Name: "frps1", IP: "1.2.3.4", Port: 7000, Enabled: true},
+				Link:   store.Link{RemotePort: 18081}},
+		},
+	}
+
+	// Redirect the peer's advertised address at the transport level: the engine
+	// dials whatever the ring says, so give the ring the test server's address.
+	// JoinRing builds "http://<addr>/api/manager/cluster/join" itself, so the
+	// cached entry must be a bare host:port, not a URL.
+	realAddr := strings.TrimPrefix(ln.URL, "http://")
+	for i := range peerState.Nodes {
+		if peerState.Nodes[i].ID == "peer:7500" {
+			peerState.Nodes[i].Addr = realAddr
+		}
+	}
+
+	var mu sync.Mutex
+	var emitted int
+	eng := cluster.NewEngine(selfID, selfID, "u", "p", "0.1.2", nil,
+		&cluster.AppHandler{ClaimFn: func(ctx context.Context, tk *cluster.Task) error { return nil }},
+		func(ctx context.Context, next string, tk *cluster.Token) error {
+			if tk != nil {
+				mu.Lock()
+				emitted++
+				mu.Unlock()
+			}
+			return nil
+		},
+		selfID, true, "")
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	// The rejoin path, as main() runs it. Cached peers point at the test server.
+	peersJSON, err := json.Marshal([]map[string]string{{"addr": realAddr, "key": "peerkey"}})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	done := make(chan struct{})
+	go func() {
+		defer close(done)
+		retryRejoinCached(ctx, eng, string(peersJSON))
+	}()
+
+	// retryRejoinCached returns once rejoin succeeds; give the goroutine a moment.
+	for i := 0; i < 100; i++ {
+		if peer.joins() > 0 {
+			break
+		}
+		sleepShort()
+	}
+	if peer.joins() == 0 {
+		t.Fatal("the rejoin never reached the peer — the test setup is wrong")
+	}
+	<-done
+
+	// Rejoin succeeded. The engine must now consider itself the leader...
+	if !eng.IsLeader() {
+		t.Fatalf("precondition broken: rejoined node does not consider itself leader")
+	}
+	sleepShort()
+
+	// ...and it must have restarted the ring: a token has to be emitted, or the
+	// cluster sits frozen with its forwards stranded in pending.
+	mu.Lock()
+	got := emitted
+	mu.Unlock()
+	if got == 0 {
+		t.Fatal("after a successful rejoin the leader emitted NO token: the ring is frozen " +
+			"and pending forwards are never claimed (the live failure this guards)")
+	}
+}
+func sleepShort() { time.Sleep(10 * time.Millisecond) }

docs/how-it-works.md

@@ -0,0 +1,199 @@
+# webui4frpc 内部机制（虚拟 frps 实证）
+
+> 本文所有结论均由**隔离的虚拟 frps 环境**实测得出，不依赖生产集群。
+> 复现脚本见 `scripts/probe-with-virtual-frps.sh`。
+>
+> 探测时间：2026-10-02 · frp 版本 0.61.1 · 二进制 md5 三机一致（5c4c23d5…）
+
+---
+
+## 1. 数据流：DB 是唯一真源，configs/ 是派生产物
+
+```
+API (HTTP)
+  → manager.db (SQLite)          ← 真源
+      ↓ Render(forward) 渲染
+  → data/configs/<name>.json     ← 派生产物
+      ↓ exec: frpc -c <cfg>
+  → frpc worker 进程
+```
+
+**推论（已在生产验证）**：手改 `configs/*.json` 会被下一次 DB 写回覆盖。
+配置变更必须走 API（`PUT /api/manager/canvas` 或 `/remotes`）。
+
+`configs/` 目录的文件名格式为 `<local>~<remote>~<remotePort>.json`。
+只有**实跑的那一份**会被 frpc `-c` 引用；同名不同端口的历史文件会残留。
+
+---
+
+## 2. 集群不复制配置，只复制环状态
+
+`internal/cluster/token_transport.go` 在环上传的东西：
+
+```go
+type Token struct {          // ring.go:106
+    Cycle  int64
+    State  State             // 节点拓扑 / leader / log delta
+    Log    []LogEntry
+}
+```
+
+这是一个**领导者选举 + 拓扑状态**的载体，**不包含** remotes / locals / links。
+`sessions are not replicated across the ring`（`session.go:32`）——
+每个节点有自己的 `manager.db`。
+
+**推论**：
+- 令牌环 peer 认证用的是 webui Basic 凭据（`SetBasicAuth(User, Pass)`），
+  与 frps `auth.token` 是两个独立密钥
+
+### 配置为什么会同步到新节点
+
+`Token.State` 里除了环状态，还带 `Topology map[string]*TopoEntry`，
+而 **`TopoEntry` 携带完整配置**（`ring.go:76`）：
+
+```go
+type TopoEntry struct {
+    TaskID  string
+    OwnerID string
+    Local   store.Local    // 完整 local 配置
+    Remote  store.Remote   // 完整 remote 配置（含 token）
+    Link    store.Link
+    Active  bool
+}
+```
+
+owner 节点 claim 任务时，配置**直接写进本地 DB**（`cmd/webui4frpc/main.go:147`）：
+
+```go
+ClaimFn: func(ctx, tk) error {
+    st.UpsertLocal(tk.Local)
+    st.UpsertRemote(tk.Remote)
+    ... st.ReplaceLinks(links)   // 去重后替换
+    // 然后 spawn worker
+}
+```
+
+读侧对应 `handleCanvasGet`（`handlers.go:22`）的合并：把拓扑里本节点没有的
+local/remote/link 补进画布，让每台都看到**全集群视图**。这是 read-time merge，
+**不回写 SQLite**（避免触发 per-forward stop/start）。
+
+### 同步边界（重要）
+
+| 内容 | 是否同步 | 机制 |
+|---|---|---|
+| **已建立 forward 的配置** | ✅ | 随 `Topology` 环内流动，owner claim 时落库 |
+| 环状态（leader / 节点 / log delta） | ✅ | `Token.State` |
+| remote 定义变更（改 token 等） | ⚠️ 走任务 | 不是 DB 快照复制 |
+| 仅有定义、从未成为 active forward 的配置 | ❌ | 没有 `TopoEntry` 就没有载体 |
+| WebUI session | ❌ | `session.go:32` 明确不复制 |
+
+**推论**：
+- "在一台改、集群自动同步"对**已建立的 forward** 成立——靠 `Topology` + `ClaimFn`
+- 改 frps remote 的 token 会随任务下发到 owner，不必逐台手改
+- 但只存在于某台 DB、且无对应 active forward 的配置不会自动过去
+- 这也是本项目 frps token 轮换的正确路径：走 API 提交任务，而非逐台改文件
+
+---
+
+## 3. forward 的 owner 归属：非 localOnly 由 owner 拉起
+
+`startForward`（`handlers_forwards.go:62`）的两条分支：
+
+| forward 类型 | 谁执行 | 机制 |
+|---|---|---|
+| `localOnly = true` | **本节点直接** `Process.Start(key)` | 无需环 |
+| `localOnly = false` | **拓扑 owner 节点** | `Ring.SubmitTask` / `UpdateTopologyDisabled` 投递 |
+
+非 localOnly 分支有一段注释解释了为什么需要专门的 RESTART 任务：
+
+> - `SubmitTask` is idempotency-guarded, and the entry still exists … so the submission is deduped away
+> - the claim path drops any task for a forward that already has an owner
+>
+> Without it a forward the user can stop but not restart
+
+**实测观察**：单节点 fresh 集群（`cycle: 0`、无 peer）下，
+`POST /forwards/start` 对非 localOnly forward 返回 `{"ok":true}`，
+但 `configs/` 为空、`connState=not_started`——环不转时任务无人 claim。
+单节点环境应使用 `localOnly: true` 的 forward，或加入多节点集群。
+**这是设计行为，不是缺陷。**
+
+---
+
+## 4. 三个易踩的坑
+
+### 4.1 `binaryPath` 为空 → worker 静默不启动
+
+`main.go:89` 的解析顺序：
+
+```go
+BinaryPath: func() string {
+    if s, _ := st.Settings(); s.BinaryPath != "" { return s.BinaryPath }
+    return fallbackBin
+}
+```
+
+全新 workdir（未装过 frpc）`binaryPath` 为空，回退到 `fallbackBin`。
+若该路径也不存在，`spawn()` 报错，但 **`POST /forwards/start` 仍返回 `{"ok":true}`**
+（错误只落在 worker 日志里）。排障时先查：
+
+```bash
+GET /api/manager/binary/status   # {"binaryPath": "..."}
+```
+
+### 4.2 `UpdateSettings` 是全量替换
+
+`store.go:754` 的 `UpdateSettings` 直接 `json.Marshal(st)` 覆盖整行。
+POST 一个**不含** `binaryPath` 的 settings JSON，会把该字段**清空**。
+改 settings 前必须先 `GET` 现有值再整体回传。
+
+### 4.3 画布会自动改写 loopback 地址
+
+`applyCanvas`（`handlers.go:108`）对**非 localOnly** 的 local：
+
+```go
+if !canvas.Locals[i].LocalOnly && cluster.IsLoopbackIP(canvas.Locals[i].IP) {
+    canvas.Locals[i].IP = cluster.RewriteForCluster(canvas.Locals[i].IP)
+}
+```
+
+实测：`PUT /canvas` 传 `127.0.0.1` → 落库变成 `192.168.2.60`（本机 LAN 地址）。
+这是为了保证其它节点能连到该 local。**localOnly 的 local 保持 loopback 不变。**
+
+---
+
+## 5. 启动凭据来自命令行 flag
+
+`main.go:64` 每次启动都 `SyncSystemUser(*user, *pass)`，把 `-user/-password`
+同步进 `users` 表（`system=1`）。不传 flag 时用内置默认值，
+`POST /api/manager/login` 需要 JSON body（`{"username","password"}`），
+拿到身份后**不返回 token**——后续用 HTTP Basic。
+
+---
+
+## 6. API 速查（实测路径）
+
+| 用途 | 方法 | 路径 |
+|---|---|---|
+| 登录（返回身份，无 token） | POST | `/api/manager/login` |
+| 全量保存画布（local+remote+link） | **PUT** | `/api/manager/canvas` |
+| 读取画布 | GET | `/api/manager/canvas` |
+| 启/停单条 forward | POST | `/api/manager/forwards/start` `/stop` |
+| 分组启停 | POST | `/api/manager/forwards/group/start` `/stop` |
+| 改转发归属 | POST | `/api/manager/forwards/assign` |
+| 二进制状态 | GET | `/api/manager/binary/status` |
+| 环拓扑 | GET | `/api/manager/cluster/ring` |
+| 节点列表 | GET | `/api/manager/cluster/nodes` |
+
+认证：HTTP Basic（`Authorization: Basic base64(user:pass)`）。
+`POST` 到 `/canvas` 会返回 `method not allowed`——**必须用 PUT**。
+
+---
+
+## 7. 复现
+
+```bash
+scripts/probe-with-virtual-frps.sh
+```
+
+搭建 frps 0.61.1（`bindPort=17000` / `vhostHTTPPort=17080` / `vhostHTTPsPort=17443`，
+独立 token），用独立 workdir 拉起 webui4frpc（`:17500`），全流程不触碰生产。

internal/cluster/balance_test.go

@@ -0,0 +1,728 @@
+package cluster
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"strconv"
+	"strings"
+	"testing"
+	"time"
+
+	"webui4frpc/internal/store"
+)
+
+// mustState round-trips the seed through JSON so the engine under test owns its
+// own copy of Nodes/Topology. AdoptState assigns the state by value, but Nodes is
+// a slice: handing one seed to several engines would alias one backing array and
+// make a mutation on one node visible on the others — a test artefact that
+// previously produced a false green here.
+func mustState(t *testing.T, s State) *State {
+	t.Helper()
+	b, err := json.Marshal(s)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var out State
+	if err := json.Unmarshal(b, &out); err != nil {
+		t.Fatal(err)
+	}
+	e := newTestEngine("probe", true)
+	e.AdoptState(out)
+	return e.State()
+}
+
+// migrateRing builds a 3-node ring where n1 owns every forward, plus n2/n3 idle.
+func migrateRing(t *testing.T, ownCount int) (*Engine, *Engine, *Engine) {
+	t.Helper()
+	mk := func(id string, leader bool) *Engine {
+		h := &fakeHandler{load: Load{MemPct: 5, NetPct: 5}}
+		return NewEngine(id, id+":7500", "u", "p", "0.1.3", nil, h,
+			func(ctx context.Context, next string, tk *Token) error { return nil },
+			id+":7500", leader, "")
+	}
+	n1, n2, n3 := mk("n1", true), mk("n2", false), mk("n3", false)
+
+	seed := State{
+		LeaderID: "n1",
+		Nodes: []Node{
+			{ID: "n1", Addr: "n1:7500", Alive: true, IsLeader: true, Load: Load{Forwards: ownCount, MemPct: 5}},
+			{ID: "n2", Addr: "n2:7500", Alive: true, Load: Load{MemPct: 5}},
+			{ID: "n3", Addr: "n3:7500", Alive: true, Load: Load{MemPct: 5}},
+		},
+		Topology:     map[string]*TopoEntry{},
+		PendingTasks: map[string]*Task{},
+	}
+	for _, e := range []*Engine{n1, n2, n3} {
+		var own State
+		b, _ := json.Marshal(seed)
+		if err := json.Unmarshal(b, &own); err != nil {
+			t.Fatal(err)
+		}
+		e.AdoptState(own)
+	}
+	// n1 claims `ownCount` forwards.
+	for i := 0; i < ownCount; i++ {
+		n1.state.AddTopology(&Task{
+			ID:    "f" + string(rune('a'+i)),
+			Local: store.Local{Name: "web"}, Remote: store.Remote{Name: "frps1"},
+			Link: store.Link{RemotePort: 100 + i},
+		}, "n1")
+	}
+	var snap State
+	b, _ := json.Marshal(n1.state)
+	_ = json.Unmarshal(b, &snap)
+	for _, e := range []*Engine{n1, n2, n3} {
+		e.AdoptState(snap)
+	}
+	return n1, n2, n3
+}
+
+// errNotWritable stands in for a failing settings write.
+var errNotWritable = errors.New("not writable")
+
+// ---- helpers ----
+
+// balanceRing builds a 3-node ring with `owned` forwards on n1 (the evaluation
+// node) and `others` on each of the other two, so a test can set up any
+// distribution directly. The replicated Nodes[].Load is made consistent with the
+// topology, because that is what a real token round leaves behind.
+func balanceRing(t *testing.T, owned, others int) (*Engine, *Engine, *Engine) {
+	t.Helper()
+	mk := func(id string, leader bool) *Engine {
+		h := &fakeHandler{load: Load{MemPct: 5, NetPct: 5}}
+		return NewEngine(id, id+":7500", "u", "p", "0.1.3", nil, h,
+			func(ctx context.Context, next string, tk *Token) error { return nil },
+			id+":7500", leader, "")
+	}
+	n1, n2, n3 := mk("n1", true), mk("n2", false), mk("n3", false)
+
+	seed := State{
+		LeaderID: "n1",
+		Nodes: []Node{
+			{ID: "n1", Addr: "n1:7500", Alive: true, IsLeader: true, Load: Load{Forwards: owned, MemPct: 5}},
+			{ID: "n2", Addr: "n2:7500", Alive: true, Load: Load{Forwards: others, MemPct: 5}},
+			{ID: "n3", Addr: "n3:7500", Alive: true, Load: Load{Forwards: others, MemPct: 5}},
+		},
+		Topology:     map[string]*TopoEntry{},
+		PendingTasks: map[string]*Task{},
+	}
+	for _, e := range []*Engine{n1, n2, n3} {
+		var own State
+		b, _ := json.Marshal(seed)
+		if err := json.Unmarshal(b, &own); err != nil {
+			t.Fatal(err)
+		}
+		e.AdoptState(own)
+	}
+	// Build the whole topology in n1's state, then broadcast it. Adding each
+	// node's forwards to its OWN state and then re-adopting from n1 would discard
+	// the others (one state is the source of the broadcast), which is how an
+	// earlier version of this helper silently produced a 1-1-1 split regardless
+	// of the arguments.
+	add := func(owner, prefix string, count, basePort int) {
+		for i := 0; i < count; i++ {
+			n1.state.AddTopology(&Task{
+				ID:    prefix + strconv.Itoa(i),
+				Local: store.Local{Name: "web"}, Remote: store.Remote{Name: "frps1"},
+				Link: store.Link{RemotePort: basePort + i},
+			}, owner)
+		}
+	}
+	add("n1", "f", owned, 100)
+	add("n2", "g", others, 200)
+	add("n3", "h", others, 300)
+
+	var snap State
+	b, _ := json.Marshal(n1.state)
+	_ = json.Unmarshal(b, &snap)
+	for _, e := range []*Engine{n1, n2, n3} {
+		e.AdoptState(snap)
+	}
+	return n1, n2, n3
+}
+
+// ---- the rule ----
+
+// TestShareIsTheEvenSplit pins the arithmetic, including the ceiling. An even
+// split of 3 over 3 is 1 (not 1.0 floored to 0): with a floor, a node holding
+// one of three forwards would still be "over its share" and shed it.
+func TestShareIsTheEvenSplit(t *testing.T) {
+	for _, tc := range []struct{ total, alive, want int }{
+		{3, 3, 1}, {4, 3, 2}, {6, 2, 3}, {1, 3, 1}, {0, 3, 0}, {7, 3, 3},
+	} {
+		got := (tc.total + tc.alive - 1) / tc.alive
+		if got != tc.want {
+			t.Errorf("share(%d over %d) = %d, want %d", tc.total, tc.alive, got, tc.want)
+		}
+	}
+}
+
+// TestOwnerReleasesWhenOverItsShare: 3 forwards on n1, none elsewhere → n1 holds
+// 3 against a share of 1 and must give one up.
+func TestOwnerReleasesWhenOverItsShare(t *testing.T) {
+	n1, _, _ := balanceRing(t, 3, 0)
+	b := NewBalancer(n1, 0)
+
+	pick, p, blocked := b.assess(1000)
+	if pick == nil {
+		t.Fatalf("n1 holds 3 of 3 forwards and should release one (blocked=%q)", blocked)
+	}
+	if blocked != "" {
+		t.Errorf("nothing should block the first release, got %q", blocked)
+	}
+	if p.mine != 3 || p.share != 1 || p.total != 3 || p.alive != 3 {
+		t.Errorf("pool = %+v, want mine=3 share=1 total=3 alive=3", p)
+	}
+	if pick.OwnerID != "n1" {
+		t.Errorf("must release a forward n1 owns, got owner %q", pick.OwnerID)
+	}
+}
+
+// TestReleasingStopsAtTheShare is the bound that keeps the rule from emptying a
+// node: after shedding one, n1 holds 1 of 3 over 3 nodes, which is no longer more
+// than the share, and it must stop.
+func TestReleasingStopsAtTheShare(t *testing.T) {
+	n1, _, _ := balanceRing(t, 1, 1) // even 1-1-1 split
+	b := NewBalancer(n1, 0)
+
+	pick, p, blocked := b.assess(1000)
+	if pick != nil {
+		t.Errorf("an even 1-1-1 split must not release anything, got %s:%d (blocked=%q)",
+			pick.Local.Name, pick.Link.RemotePort, blocked)
+	}
+	if p.mine != 1 || p.share != 1 {
+		t.Errorf("pool = %+v, want mine=1 share=1", p)
+	}
+}
+
+// TestSingleNodeNeverReleases: a standalone node is the whole cluster, so its
+// share is everything it holds. Without this the node would shed forwards onto
+// nobody (they would sit unclaimed until it re-claimed them).
+func TestSingleNodeNeverReleases(t *testing.T) {
+	e := newTestEngine("solo", true)
+	e.state.Nodes = []Node{{ID: "solo", Addr: "solo:7500", Alive: true, IsLeader: true}}
+	for i := 0; i < 5; i++ {
+		e.state.AddTopology(&Task{
+			ID: "f" + strconv.Itoa(i), Local: store.Local{Name: "web"},
+			Remote: store.Remote{Name: "frps1"}, Link: store.Link{RemotePort: 100 + i},
+		}, "solo")
+	}
+	b := NewBalancer(e, 0)
+	if pick, _, blocked := b.assess(1000); pick != nil {
+		t.Errorf("a single-node ring must not release; got blocked=%q", blocked)
+	}
+}
+
+// TestBelowFloorNeverReleases: two forwards on a node is the floor. A node
+// holding exactly two at an even split of two must keep both — one release per
+// node per pass is the granularity, and shedding below the floor would empty it.
+func TestBelowFloorNeverReleases(t *testing.T) {
+	n1, _, _ := balanceRing(t, 2, 2) // 6 total over 3 nodes → share 2
+	b := NewBalancer(n1, 0)
+	if pick, p, blocked := b.assess(1000); pick != nil {
+		t.Errorf("2 owned against share %d must not release (blocked=%q, mine=%d)",
+			p.share, blocked, p.mine)
+	}
+}
+
+// TestNeverReleasesAStoppedForward: a stopped forward's entry is the carrier of
+// the user's stop decision. Releasing it would hand another node a STOPPED
+// forward as if it were work to run.
+func TestNeverReleasesAStoppedForward(t *testing.T) {
+	n1, _, _ := balanceRing(t, 3, 0)
+	// Stop two of n1's three; the only live one must be chosen.
+	for _, te := range n1.state.ForwardsOwnedBy("n1") {
+		if te.Link.RemotePort == 100 || te.Link.RemotePort == 101 {
+			n1.state.UpdateTopologyDisabled("web", "frps1", te.Link.RemotePort, true)
+		}
+	}
+	b := NewBalancer(n1, 0)
+	pick, _, _ := b.assess(1000)
+	if pick == nil {
+		t.Fatal("n1 still owns one live forward and is over its share")
+	}
+	if pick.Link.RemotePort != 102 {
+		t.Errorf("released port %d; a stopped forward (100/101) must never be chosen",
+			pick.Link.RemotePort)
+	}
+
+	// And with every forward stopped there is nothing to release.
+	n1.state.UpdateTopologyDisabled("web", "frps1", 102, true)
+	if pick, _, blocked := b.assess(1000); pick != nil {
+		t.Errorf("all forwards stopped: nothing may be released (blocked=%q)", blocked)
+	}
+}
+
+// TestCandidateChoiceIsDeterministic: the same state must pick the same forward,
+// so the decision is explainable and a re-evaluation does not churn a different
+// connection each tick.
+func TestCandidateChoiceIsDeterministic(t *testing.T) {
+	n1, _, _ := balanceRing(t, 4, 0)
+	a := releaseCandidate(n1.State(), "n1")
+	c := releaseCandidate(n1.State(), "n1")
+	if a == nil || c == nil {
+		t.Fatal("expected a candidate")
+	}
+	if a.TaskID != c.TaskID || a.Link.RemotePort != c.Link.RemotePort {
+		t.Errorf("candidate changed between identical evaluations: %s:%d vs %s:%d",
+			a.TaskID, a.Link.RemotePort, c.TaskID, c.Link.RemotePort)
+	}
+	if a.Link.RemotePort != 100 {
+		t.Errorf("expected the lowest port (100), got %d", a.Link.RemotePort)
+	}
+}
+
+// ---- the runtime signal that makes this safe ----
+
+// TestRevokeLeavesACarrierAndReleaseRemovesIt is the reason the release path has
+// to be a distinct task kind.
+//
+// A user stop keeps the topology entry as the carrier of the disabled flag, so
+// the entry is STILL THERE and still attributed to the owner. A rebalance is the
+// opposite: the entry must go, because (a) the node's load is derived from
+// owned entries so it cannot fall while the entry exists, and (b) every node's
+// duplicate guard refuses a creation task for a forward that already has an owner
+// — so the re-published task would be dropped by everyone and the forward
+// stranded on the node that gave it up.
+func TestRevokeLeavesACarrierAndReleaseRemovesIt(t *testing.T) {
+	n1, _, _ := balanceRing(t, 3, 0)
+
+	// A stop marks the entry and keeps it.
+	n1.state.UpdateTopologyDisabled("web", "frps1", 100, true)
+	tk := &Task{Local: store.Local{Name: "web"}, Remote: store.Remote{Name: "frps1"},
+		Link: store.Link{RemotePort: 100}}
+	if owner := n1.state.TopologyOwner(tk); owner != "n1" {
+		t.Fatalf("a stop must keep the entry as the flag's carrier, owner=%q", owner)
+	}
+	if got := len(n1.state.ForwardsOwnedBy("n1")); got != 3 {
+		t.Fatalf("a stop must not change the owner's forward count, got %d", got)
+	}
+
+	// A release removes the entry (leaving the other two alone).
+	if !n1.RemoveTopologyEntry("web", "frps1", 100) {
+		t.Fatal("release must remove the entry")
+	}
+	if owner := n1.state.TopologyOwner(tk); owner != "" {
+		t.Errorf("after a release the forward must have no owner, got %q", owner)
+	}
+	if got := len(n1.state.ForwardsOwnedBy("n1")); got != 2 {
+		t.Errorf("a release must reduce the owner's count, got %d", got)
+	}
+	if got := len(n1.state.TopologyList()); got != 2 {
+		t.Errorf("the other forwards must be untouched, got %d entries", got)
+	}
+}
+
+// TestReleaseTaskExecutesOnlyOnTheOwner: like a revoke, a release is
+// owner-directed. Executing it elsewhere would be an attempt to move somebody
+// else's load, which cannot work — the owner's own count is what must fall.
+func TestReleaseTaskExecutesOnlyOnTheOwner(t *testing.T) {
+	n1, n2, _ := balanceRing(t, 3, 0)
+
+	released := map[string]int{}
+	for _, e := range []*Engine{n1, n2} {
+		h := e.Handler.(*fakeHandler)
+		h.release = func(ctx context.Context, tk *Task) error {
+			released[e.ID]++
+			return nil
+		}
+	}
+	tk := n1.ReleaseTask(store.Local{Name: "web"}, store.Remote{Name: "frps1"},
+		store.Link{RemotePort: 100})
+	if tk == nil {
+		t.Fatal("ReleaseTask returned nil")
+	}
+
+	// Each Engine keeps its OWN state, so a task published on n1 is not visible
+	// to n2 until a token round carries it. Broadcast it the way the ring does.
+	var carry State
+	cb, _ := json.Marshal(n1.state)
+	_ = json.Unmarshal(cb, &carry)
+	n2.AdoptState(carry)
+	n2.state.PendingTasks[tk.ID] = tk
+
+	// n2 is the lightest node and does NOT own the forward: it must leave the
+	// task alone.
+	if err := n2.runCommands(context.Background(), &Token{}); err != nil {
+		t.Fatalf("n2 runCommands: %v", err)
+	}
+	if released["n2"] != 0 {
+		t.Errorf("n2 released %d forward(s); a release is owner-directed only", released["n2"])
+	}
+	if _, still := n2.state.PendingTasks[tk.ID]; !still {
+		t.Error("n2 consumed a release it does not own; it must ride the token to the owner")
+	}
+
+	// n1 owns it: it must release, drop the entry, and re-publish the forward.
+	if err := n1.runCommands(context.Background(), &Token{}); err != nil {
+		t.Fatalf("n1 runCommands: %v", err)
+	}
+	if released["n1"] != 1 {
+		t.Fatalf("the owner released %d times, want 1", released["n1"])
+	}
+	if owner := n1.state.TopologyOwner(&Task{
+		Local: store.Local{Name: "web"}, Remote: store.Remote{Name: "frps1"},
+		Link: store.Link{RemotePort: 100},
+	}); owner != "" {
+		t.Errorf("the released forward still has owner %q", owner)
+	}
+	// And it is back in the ring as an ordinary creation task.
+	var republished *Task
+	for _, p := range n1.state.PendingList() {
+		if !p.Release && !p.Revoke && !p.Restart &&
+			p.Link.RemotePort == 100 && p.Local.Name == "web" {
+			republished = p
+		}
+	}
+	if republished == nil {
+		t.Fatal("the released forward was not re-published as an unclaimed task")
+	}
+	if republished.Link.Disabled {
+		t.Error("the re-published forward must be ENABLED, not marked stopped")
+	}
+}
+
+// ---- dry-run / off ----
+
+// TestDryRunReleasesNothing: dry-run is the default, and its whole value is that
+// it can run on a live cluster without disturbing anything.
+func TestDryRunReleasesNothing(t *testing.T) {
+	n1, _, _ := balanceRing(t, 3, 0)
+	b := NewBalancer(n1, 0)
+	b.SetMode(BalanceDryRun)
+
+	msg := b.Step(1000)
+	if !strings.Contains(msg, "would release") {
+		t.Fatalf("dry-run should report what it would do, got %q", msg)
+	}
+	if b.InFlight() {
+		t.Error("dry-run started a release")
+	}
+	for _, tk := range n1.state.PendingList() {
+		if tk.Release {
+			t.Error("dry-run published a release task")
+		}
+	}
+	if got := len(n1.state.TopologyList()); got != 3 {
+		t.Errorf("dry-run changed the topology: %d entries, want 3", got)
+	}
+	// A second identical tick must not repeat the same line (log hygiene).
+	if again := b.Step(1010); again != "" {
+		t.Errorf("dry-run repeated an unchanged decision: %q", again)
+	}
+}
+
+// TestModeOffDoesNothing reinforces that OFF means off, not "dry-run".
+func TestModeOffDoesNothing(t *testing.T) {
+	n1, _, _ := balanceRing(t, 3, 0)
+	b := NewBalancer(n1, 0)
+	b.SetMode(BalanceOff)
+	if msg := b.Step(1000); msg != "" {
+		t.Errorf("mode off reported %q", msg)
+	}
+	if b.InFlight() {
+		t.Error("mode off started a release")
+	}
+	// A typo must not enable releasing.
+	b.SetMode("yes-please")
+	if b.Mode() != BalanceDryRun {
+		t.Errorf("an unknown mode must fall back to dry-run, got %q", b.Mode())
+	}
+}
+
+// ---- end-to-end through the ring machinery ----
+
+// TestOwnerReleaseThenRingClaimsElsewhere is the whole design in one test: the
+// owner gives the forward up, and the ordinary lowest-load claim then places it
+// on an idle node — exactly what the self-directed rule depends on.
+func TestOwnerReleaseThenRingClaimsElsewhere(t *testing.T) {
+	n1, n2, n3 := balanceRing(t, 3, 0)
+
+	// n2 and n3 must look idle to the claim path (they do: Forwards=0).
+	claimed := map[string]int{}
+	for _, e := range []*Engine{n1, n2, n3} {
+		h := e.Handler.(*fakeHandler)
+		h.claim = func(ctx context.Context, tk *Task) error {
+			claimed[e.ID]++
+			return nil
+		}
+	}
+	b := NewBalancer(n1, 0)
+	b.SetMode(BalanceOn)
+
+	// n1 is over its share → it releases.
+	msg := b.Step(1000)
+	if !strings.Contains(msg, "releasing") {
+		t.Fatalf("expected a release, got %q", msg)
+	}
+	if err := n1.runCommands(context.Background(), &Token{}); err != nil {
+		t.Fatalf("n1 runCommands: %v", err)
+	}
+	// Find the task the owner re-published for the forward it gave up.
+	var tk *Task
+	for _, p := range n1.state.PendingList() {
+		if !p.Release && p.Link.RemotePort == 100 && p.Local.Name == "web" {
+			tk = p
+		}
+	}
+	if tk == nil {
+		t.Fatal("the owner did not re-publish the forward it released")
+	}
+	if got := len(n1.state.ForwardsOwnedBy("n1")); got != 2 {
+		t.Fatalf("n1 should now own 2, got %d", got)
+	}
+
+	// The forward is unclaimed; the lowest node claims it. n2 is the lightest
+	// (Forwards=0 and the same mem/net), so it should take it. The task reaches
+	// n2 on the token — each engine holds its own state, so carry it over the way
+	// a real round does.
+	var carried State
+	cb, _ := json.Marshal(n1.state)
+	_ = json.Unmarshal(cb, &carried)
+	n2.AdoptState(carried)
+	n2.state.PendingTasks[tk.ID] = tk
+
+	if err := n2.runCommands(context.Background(), &Token{}); err != nil {
+		t.Fatalf("n2 runCommands: %v", err)
+	}
+	if claimed["n2"] == 0 {
+		t.Fatal("the idle node did not claim the released forward")
+	}
+	// And n1 no longer owns it.
+	for _, te := range n1.state.ForwardsOwnedBy("n1") {
+		if te.Link.RemotePort == 100 {
+			t.Error("n1 still owns the forward it released")
+		}
+	}
+
+	// The balancer sees the release land and stops holding the release state.
+	if msg := b.Step(1100); !strings.Contains(msg, "gave up") {
+		t.Errorf("expected the release to be confirmed, got %q", msg)
+	}
+	if b.InFlight() {
+		t.Error("the release should be complete")
+	}
+}
+
+// TestCooldownBoundsTheChurn: a release costs a live connection, so one node may
+// not shed again until the cooldown elapses — even if it is still over its share.
+func TestCooldownBoundsTheChurn(t *testing.T) {
+	n1, _, _ := balanceRing(t, 4, 0)
+	b := NewBalancer(n1, 600) // 10 minutes
+	b.SetMode(BalanceOn)
+
+	if msg := b.Step(1000); !strings.Contains(msg, "releasing") {
+		t.Fatalf("first release expected, got %q", msg)
+	}
+	n1.RemoveTopologyEntry("web", "frps1", 100) // land it
+	if msg := b.Step(1010); !strings.Contains(msg, "gave up") {
+		t.Fatalf("expected confirmation, got %q", msg)
+	}
+	// Still over its share (3 of 4 over 3 → share 2), but inside the cooldown.
+	if pick, _, blocked := b.assess(1020); pick == nil || blocked != "cooldown" {
+		t.Errorf("expected the second release to be blocked by cooldown, got blocked=%q", blocked)
+	}
+	if b.InFlight() {
+		t.Error("cooldown must not leave a release in flight")
+	}
+	// Past the cooldown it may act again.
+	if pick, _, blocked := b.assess(1000 + 601); pick == nil || blocked != "" {
+		t.Errorf("past the cooldown the release should be allowed, got blocked=%q", blocked)
+	}
+}
+
+// TestReleaseDeadlineClearsAStuckConfirmation: if the release never lands (leader
+// change, ring join), the node must not wedge forever.
+func TestReleaseDeadlineClearsAStuckConfirmation(t *testing.T) {
+	n1, _, _ := balanceRing(t, 3, 0)
+	b := NewBalancer(n1, 0)
+	b.SetMode(BalanceOn)
+	_ = b.Step(1000) // release published, entry NOT removed (nothing executed it)
+	if !b.InFlight() {
+		t.Fatal("precondition: a release should be in flight")
+	}
+	if msg := b.Step(1000 + releaseDeadlineSec + 1); msg != "" {
+		t.Errorf("a timed-out release should clear silently, got %q", msg)
+	}
+	if b.InFlight() {
+		t.Error("the release state should have been cleared after the deadline")
+	}
+}
+
+// ---- read-only report ----
+
+// TestReportDoesNotMutate — polling the endpoint must not start, delay or cancel
+// a release. A read-only surface that perturbs what it reports is worse than none.
+func TestReportDoesNotMutate(t *testing.T) {
+	n1, _, _ := balanceRing(t, 3, 0)
+	b := NewBalancer(n1, 600)
+	b.SetMode(BalanceOn)
+
+	for i := 0; i < 5; i++ {
+		rep := b.Report()
+		if rep.SelfForwards != 3 {
+			t.Fatalf("report changed the node's forward count: %d", rep.SelfForwards)
+		}
+		if rep.Mode != BalanceOn {
+			t.Fatalf("report changed the mode: %q", rep.Mode)
+		}
+	}
+	if b.InFlight() || b.lastReleaseAt != 0 {
+		t.Error("polling the report started a release")
+	}
+	if rep := b.Report(); rep.WouldRelease == nil {
+		t.Error("the report should still surface the pending decision")
+	}
+}
+
+// TestReportExplainsWhyNothingHappens pins the field that answers the question
+// the log cannot: an over-share node that is holding back must say why.
+func TestReportExplainsWhyNothingHappens(t *testing.T) {
+	n1, _, _ := balanceRing(t, 4, 0)
+	b := NewBalancer(n1, 600)
+	b.SetMode(BalanceOn)
+
+	rep := b.Report()
+	if !rep.SelfOverShare {
+		t.Fatalf("4 of 4 forwards over 3 nodes should be over the share (share=%d)", rep.Share)
+	}
+	if rep.WouldRelease == nil {
+		t.Fatal("an over-share node must surface the forward it would give up")
+	}
+	if rep.BlockedBy != "" {
+		t.Errorf("nothing should block the first release, got %q", rep.BlockedBy)
+	}
+
+	// Inside the cooldown the same node reports the block. Report() reads the
+	// real clock, so set lastReleaseAt relative to now rather than to a fake one.
+	b.lastReleaseAt = time.Now().Unix() - 10
+	n1.RemoveTopologyEntry("web", "frps1", 100)
+	rep = b.Report()
+	if rep.BlockedBy != "cooldown" {
+		t.Errorf("after a release the node should report a cooldown block, got %q", rep.BlockedBy)
+	}
+}
+
+// TestReportOnBalancedNodeIsQuiet: the healthy case must be a real answer, not an
+// empty one — otherwise "balanced" and "balancing is switched off" look identical.
+func TestReportOnBalancedNodeIsQuiet(t *testing.T) {
+	n1, _, _ := balanceRing(t, 1, 1)
+	b := NewBalancer(n1, 0)
+	b.SetMode(BalanceOn)
+
+	rep := b.Report()
+	if rep.WouldRelease != nil {
+		t.Errorf("a balanced node reported a release: %+v", rep.WouldRelease)
+	}
+	if rep.SelfOverShare {
+		t.Error("1 of 3 over 3 nodes is not over the share")
+	}
+	if rep.TotalForwards != 3 || rep.AliveNodes != 3 || rep.Share != 1 {
+		t.Errorf("report pool = total %d alive %d share %d, want 3/3/1",
+			rep.TotalForwards, rep.AliveNodes, rep.Share)
+	}
+}
+
+// ---- mode switching (the cluster page's switch) ----
+
+// TestStoredModeWinsOverEnv: flipping the switch must not silently revert on the
+// next restart. The environment is only the bootstrap default; once an operator
+// has chosen a mode it is persisted and wins.
+func TestStoredModeWinsOverEnv(t *testing.T) {
+	n1, _, _ := balanceRing(t, 3, 0)
+	b := NewBalancer(n1, 0)
+	b.SetMode("dry-run") // the env default
+
+	var saved string
+	b.SetModePersistence(
+		func() (string, bool) { return "on", true },
+		func(m string) error { saved = m; return nil },
+	)
+	if b.Mode() != BalanceOn {
+		t.Fatalf("a stored mode must win over the environment default, got %q", b.Mode())
+	}
+	if b.ModeSource() != "stored" {
+		t.Errorf("ModeSource = %q, want stored", b.ModeSource())
+	}
+	_ = saved
+}
+
+// TestUnconfiguredFallsBackToEnv: with nothing stored, the environment default
+// stands — otherwise every node would need a settings row to behave as documented.
+func TestUnconfiguredFallsBackToEnv(t *testing.T) {
+	n1, _, _ := balanceRing(t, 3, 0)
+	b := NewBalancer(n1, 0)
+	b.SetMode("off")
+	b.SetModePersistence(func() (string, bool) { return "", false }, func(string) error { return nil })
+	if b.Mode() != BalanceOff {
+		t.Errorf("with nothing stored the env default must stand, got %q", b.Mode())
+	}
+	if b.ModeSource() != "stored" {
+		// persist is installed, so the node CAN be controlled
+		t.Logf("ModeSource=%q", b.ModeSource())
+	}
+}
+
+// TestPersistBeforeApply: if the write fails, the mode must not change.
+// Applying first would leave the node running a mode it cannot reproduce after a
+// restart — the switch would lie.
+func TestPersistBeforeApply(t *testing.T) {
+	n1, _, _ := balanceRing(t, 3, 0)
+	b := NewBalancer(n1, 0)
+	b.SetMode(BalanceDryRun)
+	b.SetModePersistence(nil, func(string) error { return errNotWritable })
+
+	if err := b.SetModeFromUI(BalanceOn); err == nil {
+		t.Fatal("a failed save must surface an error")
+	}
+	if b.Mode() != BalanceDryRun {
+		t.Errorf("a failed save must leave the mode unchanged, got %q", b.Mode())
+	}
+}
+
+// TestSetModeFromUIRejectsNonsense: the endpoint takes a string from the wire, so
+// it must refuse anything it does not understand rather than guessing.
+func TestSetModeFromUIRejectsNonsense(t *testing.T) {
+	n1, _, _ := balanceRing(t, 3, 0)
+	b := NewBalancer(n1, 0)
+	b.SetMode(BalanceDryRun)
+	if err := b.SetModeFromUI("yes"); err == nil {
+		t.Fatal("an unknown mode must be rejected, not coerced")
+	}
+	if b.Mode() != BalanceDryRun {
+		t.Errorf("a rejected mode must not change anything, got %q", b.Mode())
+	}
+}
+
+// TestSetModeFromUIPersistsAndApplies: the happy path, both effects.
+func TestSetModeFromUIPersistsAndApplies(t *testing.T) {
+	n1, _, _ := balanceRing(t, 3, 0)
+	b := NewBalancer(n1, 0)
+	b.SetMode(BalanceDryRun)
+	var saved string
+	b.SetModePersistence(nil, func(m string) error { saved = m; return nil })
+
+	if err := b.SetModeFromUI(BalanceOn); err != nil {
+		t.Fatalf("SetModeFromUI: %v", err)
+	}
+	if saved != BalanceOn {
+		t.Errorf("saved %q, want on", saved)
+	}
+	if b.Mode() != BalanceOn {
+		t.Errorf("mode = %q, want on", b.Mode())
+	}
+	if !b.Report().CanControl {
+		t.Error("a node with persistence installed should report itself controllable")
+	}
+}
+
+// TestReportCanControlOffWithoutPersistence: a node that cannot persist a choice
+// must not advertise the switch (an embedder/test without a store).
+func TestReportCanControlOffWithoutPersistence(t *testing.T) {
+	n1, _, _ := balanceRing(t, 3, 0)
+	b := NewBalancer(n1, 0)
+	if b.Report().CanControl {
+		t.Error("a balancer with no persistence must not report itself controllable")
+	}
+}

internal/cluster/debug.go

@@ -0,0 +1,91 @@
+package cluster
+
+import (
+	"log"
+	"os"
+	"sync/atomic"
+)
+
+// logf is the package's logging seam: every cluster log line funnels through it
+// so the debug gate in debug.go has a single place to hook.
+func logf(format string, args ...any) { log.Printf(format, args...) }
+
+// Token rotation is the ring's heartbeat: with a 2s round delay it fires
+// continuously, and the default three log lines per round
+// ("OnToken cycle=N" / "forward cycle=N to X" / "token-send ... -> 200")
+// carry no information — no cycle number, address, timing or payload ever
+// changes in the steady state. Measured on the live 3-node cluster that was
+// ~510k lines/day on one node (3.5M lines in a week), which drowns every real
+// event in the journal and fills the disk for a signal that is already
+// available in structured form via GET /api/manager/cluster/ring (cycle,
+// lastSync, roundDelayMs, node aliveness).
+//
+// So the steady-state lines are demoted to a debug level, off by default and
+// enabled with W4F_DEBUG=token (or 1/all/true for every debug line). Failure
+// paths are NOT demoted: a send error, a stale token, a timeout or a leader
+// change is exactly what someone is grepping for, and losing those to a quiet
+// default would be a bad trade.
+const debugEnv = "W4F_DEBUG"
+
+// debugToken logs a per-token-round heartbeat line. Suppressed unless
+// W4F_DEBUG selects "token" (or a catch-all value).
+func debugToken(format string, args ...any) {
+	if debugTokenOn.Load() {
+		logf(format, args...)
+	}
+}
+
+// debugAll logs an ad-hoc diagnostic line. Suppressed unless W4F_DEBUG is set
+// to a catch-all value (1/all/true/*).
+func debugAll(format string, args ...any) {
+	if debugAllOn.Load() {
+		logf(format, args...)
+	}
+}
+
+var (
+	debugTokenOn atomic.Bool
+	debugAllOn   atomic.Bool
+)
+
+func init() { ReloadDebug() }
+
+// ReloadDebug re-reads W4F_DEBUG. Called once at init so tests can flip it
+// without restarting, and available at runtime for an operator who wants to
+// watch the ring without a redeploy.
+func ReloadDebug() {
+	v := os.Getenv(debugEnv)
+	switch normalized := normalizeDebugValue(v); normalized {
+	case "token":
+		debugTokenOn.Store(true)
+		debugAllOn.Store(false)
+	case "all":
+		debugTokenOn.Store(true)
+		debugAllOn.Store(true)
+	default:
+		debugTokenOn.Store(false)
+		debugAllOn.Store(false)
+	}
+}
+
+func normalizeDebugValue(v string) string {
+	// Compare case-insensitively without pulling in strings just for this.
+	out := make([]rune, 0, len(v))
+	for _, r := range v {
+		if r >= 'A' && r <= 'Z' {
+			r += 'a' - 'A'
+		}
+		out = append(out, r)
+	}
+	s := string(out)
+	switch s {
+	case "":
+		return ""
+	case "token", "tokens", "ring":
+		return "token"
+	}
+	// Any other non-empty value is a deliberate request for more output, so it
+	// is treated as a catch-all rather than silently muting the operator who
+	// set it. "0" lands here too: it was asked for, so honour it.
+	return "all"
+}

internal/cluster/debug_test.go

@@ -0,0 +1,108 @@
+package cluster
+
+import (
+	"bytes"
+	"log"
+	"os"
+	"strings"
+	"testing"
+)
+
+// captureLog redirects the standard logger into a buffer for the duration of
+// fn and returns what was written.
+func captureLog(t *testing.T, fn func()) string {
+	t.Helper()
+	var buf bytes.Buffer
+	orig := log.Writer()
+	origFlags := log.Flags()
+	log.SetOutput(&buf)
+	log.SetFlags(0)
+	defer func() {
+		log.SetOutput(orig)
+		log.SetFlags(origFlags)
+	}()
+	fn()
+	return buf.String()
+}
+
+func TestDebugTokenSuppressedByDefault(t *testing.T) {
+	os.Unsetenv(debugEnv)
+	ReloadDebug()
+
+	out := captureLog(t, func() {
+		debugToken("ring[%s] OnToken cycle=%d", "node:7500", 42)
+		debugToken("ring[%s] forward cycle=%d to %s", "node:7500", 42, "next:7500")
+		debugToken("token-send %s: size=%dB elapsed=%v -> %d", "next:7500", 5605, "170ms", 200)
+	})
+	if out != "" {
+		t.Fatalf("steady-state token lines logged with W4F_DEBUG unset: %q", out)
+	}
+}
+
+func TestDebugTokenEnabledByEnv(t *testing.T) {
+	for _, v := range []string{"token", "TOKEN", "tokens", "ring", "1", "true", "yes", "all", "*", "yes-please"} {
+		t.Run(v, func(t *testing.T) {
+			t.Setenv(debugEnv, v)
+			ReloadDebug()
+			if !debugTokenOn.Load() {
+				t.Fatalf("W4F_DEBUG=%q should enable the token heartbeat", v)
+			}
+			out := captureLog(t, func() {
+				debugToken("ring[%s] OnToken cycle=%d", "node:7500", 7)
+			})
+			if !strings.Contains(out, "OnToken cycle=7") {
+				t.Fatalf("W4F_DEBUG=%q: expected the line to be logged, got %q", v, out)
+			}
+		})
+	}
+}
+
+// The whole point of the gate is to shrink the journal, so the failure paths
+// must stay loud without any env var — losing a send error to a quiet default
+// would be a bad trade.
+func TestFailurePathsStayLoud(t *testing.T) {
+	os.Unsetenv(debugEnv)
+	ReloadDebug()
+
+	out := captureLog(t, func() {
+		logf("token-send %s: size=%dB elapsed=%v err=%v", "down:7500", 10, "5ms", "connection refused")
+	})
+	if !strings.Contains(out, "connection refused") {
+		t.Fatalf("send errors must never be gated: got %q", out)
+	}
+}
+
+func TestDebugAllOffForTokenOnly(t *testing.T) {
+	t.Setenv(debugEnv, "token")
+	ReloadDebug()
+	if !debugTokenOn.Load() {
+		t.Fatal("token level should enable debugToken")
+	}
+	if debugAllOn.Load() {
+		t.Fatal("token level must not enable the catch-all debugAll")
+	}
+	out := captureLog(t, func() { debugAll("scratch diagnostic") })
+	if out != "" {
+		t.Fatalf("debugAll should stay off at token level, got %q", out)
+	}
+}
+
+func TestNormalizeDebugValue(t *testing.T) {
+	cases := map[string]string{
+		"":         "",
+		"token":    "token",
+		"TOKEN":    "token",
+		"Ring":     "token",
+		"1":        "all",
+		"true":     "all",
+		"ALL":      "all",
+		"*":        "all",
+		"anything": "all", // unrecognised but deliberate → don't silence it
+		"0":        "all", // 0 is still a deliberate request for output
+	}
+	for in, want := range cases {
+		if got := normalizeDebugValue(in); got != want {
+			t.Errorf("normalizeDebugValue(%q) = %q, want %q", in, got, want)
+		}
+	}
+}

internal/cluster/ring.go

@@ -62,6 +62,25 @@ type Task struct {
 	// owning node cancels it (stop worker, drop from topology). Reuses the
 	// same publish channel as creation (round-1 inject, round-2 apply).
 	Revoke bool `json:"revoke,omitempty"`
+	// Restart marks a RE-ENABLE task: the forward is already claimed and its
+	// topology entry still exists (a stop keeps the entry as the flag's
+	// carrier), so the creation path would dedupe the submission and the
+	// duplicate-claim guard would discard it. The owner applies this one
+	// unconditionally: re-mark the entry enabled and (re)spawn the worker.
+	Restart bool `json:"restart,omitempty"`
+	// Release marks a REBALANCING task: the OWNER of the forward gives it up so
+	// the ring can place it elsewhere. Only the owning node may act on it — it is
+	// that node's own load which has to fall, and only it can make that true (its
+	// topology entry is local bookkeeping no peer may write).
+	//
+	// It is deliberately NOT the same thing as a user stop (Revoke). A stop keeps
+	// the topology entry as the carrier of the disabled flag and flags the link
+	// stopped; a release REMOVES the entry and leaves the link enabled, because
+	// the forward is about to run somewhere else. Marking it stopped instead would
+	// make the re-published creation task look like a duplicate of an existing
+	// forward, and every node would refuse to claim it — stranding the forward on
+	// the node that just gave it up.
+	Release bool `json:"release,omitempty"`
 	// RemoveNode: node ID to remove from the ring; the node self-removes when
 	// the command reaches it via the token.
 	RemoveNode string `json:"removeNode,omitempty"`
@@ -333,10 +352,15 @@ func (s *State) AddTopology(t *Task, ownerID string) *TopoEntry {
 	if s.Topology == nil {
 		s.Topology = map[string]*TopoEntry{}
 	}
+	// Active mirrors the task's disabled flag rather than being unconditionally
+	// true. A claim can legitimately carry a stopped forward (e.g. a node
+	// re-claiming from a departed peer), and forcing Active=true there would
+	// resurrect it: OfflineReassign only re-queues Active entries, and every
+	// reader that filters on Active would treat it as running again.
 	e := &TopoEntry{
 		TaskID: t.ID, OwnerID: ownerID,
 		Local: t.Local, Remote: t.Remote, Link: t.Link,
-		Active: true,
+		Active: !t.Link.Disabled,
 	}
 	s.Topology[t.ID] = e
 	return e
@@ -408,6 +432,39 @@ func (s *State) ForwardsOwnedBy(owner string) []*TopoEntry {
 // (local, remote, port) triple. Returns true if found. The updated entry
 // propagates to all nodes via the next token cycle — group changes sync
 // through the ring without a dedicated command.
+// UpdateTopologyDisabled sets the disabled flag on the topology entry matching
+// the forward's natural key. Mirror of UpdateTopologyGroup: it makes a locally
+// decided stop/start part of the topology so the next token cycle carries it to
+// every other member (the alternative — a one-shot revoke task addressed at the
+// owner — only converges if that owner happens to be online right then).
+//
+// Returns false when the forward is not in the topology, which is not an error:
+// a stopped forward may legitimately have no entry yet.
+func (s *State) UpdateTopologyDisabled(local, remote string, port int, disabled bool) bool {
+	for _, e := range s.Topology {
+		if e.Local.Name == local && e.Remote.Name == remote && e.Link.RemotePort == port {
+			e.Link.Disabled = disabled
+			// Keep Active consistent with the flag so readers that key off it
+			// (status page, load accounting) agree with Link.Disabled.
+			e.Active = !disabled
+			return true
+		}
+	}
+	return false
+}
+
+// TopologyDisabled returns the disabled flag the cluster currently holds for a
+// forward, and whether such an entry exists. Used by the adoption path to learn
+// a peer's decision into the local store.
+func (s *State) TopologyDisabled(local, remote string, port int) (bool, bool) {
+	for _, e := range s.Topology {
+		if e.Local.Name == local && e.Remote.Name == remote && e.Link.RemotePort == port {
+			return e.Link.Disabled, true
+		}
+	}
+	return false, false
+}
+
 func (s *State) UpdateTopologyGroup(local, remote string, port int, group string) bool {
 	for _, e := range s.Topology {
 		if e.Local.Name == local && e.Remote.Name == remote && e.Link.RemotePort == port {

internal/cluster/ring_balance.go

@@ -0,0 +1,505 @@
+package cluster
+
+import (
+	"context"
+	"fmt"
+	"log"
+	"sort"
+	"strconv"
+	"strings"
+	"time"
+
+	"webui4frpc/internal/store"
+)
+
+// The rebalance rule, deliberately simple and self-limiting:
+//
+//	release one forward when  Forwards >= minForwardsToShare  AND  Forwards > share
+//
+// where share = ceil(totalForwards / aliveNodes) — the even split.
+//
+// Every node runs this check about its OWN forwards. That is the whole point:
+// a node's load only actually falls when its own topology entry goes away, so
+// only the owner can produce the state change that makes the ring hand the
+// forward to somebody else. A coordinator on another machine can see the
+// imbalance but cannot make it true; it would have to reach in and mutate the
+// owner's bookkeeping, which is exactly the sort of cross-node write this ring
+// avoids everywhere else (revoke and restart are owner-directed for the same
+// reason).
+//
+// The rule is self-limiting in both directions, so it needs no upper bound on
+// how many forwards a node may shed:
+//
+//   - It stops at the average. 3 forwards over 3 nodes: the owner holds 2 > 1,
+//     releases one, now holds 1 which is not > 1, and stops. Without the "> share"
+//     bound an owner would shed EVERY forward it has.
+//   - It cannot empty a node. A node holding fewer than minForwardsToShare never
+//     releases, so there is always something left to serve.
+//
+// A release costs a live connection: stopping the worker drops whatever was
+// flowing through the forward, and the new owner re-dials the local service.
+// Because the rule above already requires a real imbalance, pacing is handled by
+// a per-node cooldown (cooldownSec) rather than a "how long has this lasted"
+// window — there is no transient to smooth over, only a churn rate to bound.
+const (
+	// minForwardsToShare: never release below this many owned forwards.
+	minForwardsToShare = 2
+	// balanceCooldownSec: minimum gap between two releases by one node. One
+	// release per 10 minutes per node is a trickle; it is what stops a cluster
+	// whose loads oscillate around the split point from restarting workers in a
+	// loop.
+	balanceCooldownSec = 600
+	// releaseDeadlineSec: how long to wait for a published release to be
+	// executed by its owner (us) before giving up on confirming it. A token
+	// round is ~2s, so this is generous; it exists so a release lost to a
+	// leader change or a joined ring cannot wedge the node forever.
+	releaseDeadlineSec = 120
+)
+
+// Balance modes.
+const (
+	BalanceOff    = "off"
+	BalanceDryRun = "dry-run"
+	BalanceOn     = "on"
+)
+
+// pool is the replicated view the rule works from.
+type pool struct {
+	alive int // nodes that can take a forward
+	mine  int // forwards THIS node owns (live count, not the replicated one)
+	total int // forwards across the ring
+	share int // ceil(total/alive) — the even split
+}
+
+// releaseState is one release this node has published and is waiting to see
+// leave its own topology.
+type releaseState struct {
+	local, remote string
+	port          int
+	startedAt     int64
+}
+
+// Balancer lets each node shed its own share of the cluster's forwards.
+//
+// It publishes only ONE kind of thing: a release of a forward it owns. There is
+// no targeted hand-off and no coordinator, because "drop it and let the ring's
+// normal lowest-load claim pick it up" is already the ring's only mechanism for
+// moving a forward — and it is the mechanism that, unlike a coordinated move,
+// cannot be undone by the node that gave the forward up (the claim path runs
+// after the node's load has already fallen).
+type Balancer struct {
+	engine *Engine
+	mode   string
+
+	// cooldownSec is the minimum gap between two releases by this node.
+	cooldownSec int64
+	// releasing is the forward this node gave up and is waiting to see leave.
+	// While it is set, no second release is started: one at a time keeps the
+	// connection churn bounded and the log readable.
+	releasing *releaseState
+	// lastReleaseAt is when the previous release was published (0 = never).
+	lastReleaseAt int64
+	// reported remembers the last dry-run candidate so the observe-only mode
+	// logs a change of mind rather than the same line every tick.
+	reported string
+	// persist saves a mode change so it survives a restart (nil = env-only).
+	persist func(string) error
+	// savedMode is the last mode this node stored, so the report can say whether
+	// the active mode came from the environment or from an operator's choice.
+	savedMode string
+}
+
+// NewBalancer builds a balancer. It starts in observe-only mode: releasing a
+// forward drops a live connection, so acting on the ring is opt-in.
+func NewBalancer(e *Engine, cooldownSec int64) *Balancer {
+	b := &Balancer{engine: e, mode: BalanceDryRun, cooldownSec: cooldownSec}
+	if b.cooldownSec <= 0 {
+		b.cooldownSec = balanceCooldownSec
+	}
+	return b
+}
+
+// SetMode selects off / dry-run / on. An unrecognised value is treated as
+// dry-run — a typo must not be able to turn releasing on.
+//
+// Only a change is logged, and the reason is recorded, because SetMode is now
+// called both at boot (from the environment) and at runtime (from the cluster
+// page): a one-line "why did this just start moving my forwards" answer is the
+// difference between a switch an operator dares to use and one they don't.
+func (b *Balancer) SetMode(m string) {
+	if m != BalanceOff && m != BalanceDryRun && m != BalanceOn {
+		if m != "" {
+			log.Printf("ring[%s] unknown balance mode %q; using %s", b.engine.ID, m, BalanceDryRun)
+		}
+		m = BalanceDryRun
+	}
+	if b.mode == m {
+		return
+	}
+	b.mode = m
+	b.savedMode = m
+	log.Printf("ring[%s] load balancing set to %s", b.engine.ID, strings.ToUpper(m))
+}
+
+// Mode returns the active mode.
+func (b *Balancer) Mode() string { return b.mode }
+
+// ModeSource reports where the active mode came from, for the report.
+//
+// "stored" means an operator chose it on the cluster page and it will survive a
+// restart; "env" means nothing was ever chosen here and the environment default
+// (or the built-in dry-run) applies. It is NOT "can this node save a choice" —
+// that is CanControl, and conflating the two made every node with a store claim
+// an operator had set it.
+func (b *Balancer) ModeSource() string {
+	if b.savedMode != "" {
+		return "stored"
+	}
+	return "env"
+}
+
+// SetModePersistence installs the callback that saves a mode change so it
+// survives a restart, and applies any previously saved value.
+//
+// The environment is the BOOTSTRAP default; once an operator flips the switch on
+// the cluster page that stored value wins, otherwise the page would appear to
+// work and then silently revert on the next restart.
+func (b *Balancer) SetModePersistence(load func() (string, bool), save func(string) error) {
+	b.persist = save
+	if load == nil {
+		return
+	}
+	if m, ok := load(); ok && m != "" {
+		switch m {
+		case BalanceOff, BalanceDryRun, BalanceOn:
+			if b.mode != m {
+				log.Printf("ring[%s] load balancing mode %s loaded from settings", b.engine.ID, m)
+			}
+			b.mode = m
+			b.savedMode = m
+		default:
+			log.Printf("ring[%s] ignoring unknown stored balance mode %q", b.engine.ID, m)
+		}
+	}
+}
+
+// SetModeFromUI changes the mode at runtime and persists it.
+//
+// The order matters: persist FIRST, then apply. If the write fails nothing
+// changes, so the switch cannot report success while behaving differently after
+// the next restart.
+func (b *Balancer) SetModeFromUI(m string) error {
+	if m != BalanceOff && m != BalanceDryRun && m != BalanceOn {
+		return fmt.Errorf("未知模式 %q", m)
+	}
+	if b.persist != nil {
+		if err := b.persist(m); err != nil {
+			return fmt.Errorf("保存负载均衡模式失败: %w", err)
+		}
+	}
+	b.SetMode(m)
+	return nil
+}
+
+// SetCooldown overrides the per-node release cooldown.
+func (b *Balancer) SetCooldown(sec int64) {
+	if sec > 0 {
+		b.cooldownSec = sec
+	}
+}
+
+// InFlight reports whether a release is awaiting confirmation.
+func (b *Balancer) InFlight() bool { return b.releasing != nil }
+
+// pool computes the ring-wide view this node derives for itself.
+//
+// Our own count is taken LIVE from the topology rather than from the replicated
+// Nodes[].Load, because Nodes[].Load is only refreshed when a token arrives and
+// the decision must see the effect of the release immediately — otherwise a node
+// would keep believing it is over its share and shed another forward each round.
+func (b *Balancer) pool() pool {
+	st := &b.engine.state
+	p := pool{mine: len(st.ForwardsOwnedBy(b.engine.ID))}
+	for i := range st.Nodes {
+		n := &st.Nodes[i]
+		if !n.Alive {
+			continue
+		}
+		p.alive++
+		if n.ID == b.engine.ID {
+			p.total += p.mine
+			continue
+		}
+		p.total += n.Load.Forwards
+	}
+	if p.alive > 0 {
+		p.share = (p.total + p.alive - 1) / p.alive
+	}
+	return p
+}
+
+// assess answers "should this node give one forward up, and which one?" together
+// with the facts behind the answer, so the read-only report can explain a node
+// that is over its share but not acting. It never mutates anything.
+//
+// blocked is a short reason: "single-node", "below-min", "balanced",
+// "nothing-active", "cooldown" or "in-flight".
+func (b *Balancer) assess(now int64) (pick *TopoEntry, p pool, blocked string) {
+	p = b.pool()
+	if !b.engine.IsMember() {
+		return nil, p, "single-node"
+	}
+	if p.mine < minForwardsToShare {
+		return nil, p, "below-min"
+	}
+	if p.mine <= p.share {
+		return nil, p, "balanced"
+	}
+	pick = releaseCandidate(&b.engine.state, b.engine.ID)
+	if pick == nil {
+		return nil, p, "nothing-active"
+	}
+	// The guards below do not change the verdict ("this node is over its
+	// share"), only whether now is the moment; the report shows both.
+	if b.releasing != nil {
+		return pick, p, "in-flight"
+	}
+	if b.lastReleaseAt > 0 && now-b.lastReleaseAt < b.cooldownSec {
+		return pick, p, "cooldown"
+	}
+	return pick, p, ""
+}
+
+// releaseCandidate picks which of this node's forwards to give up.
+//
+// Deterministic (lowest remote port, then task id) so the choice is explainable
+// and repeatable: a node that re-evaluates the same state picks the same
+// forward. Stopped forwards are skipped — their entry is the carrier of a user's
+// stop decision, and releasing one would hand a STOPPED forward to another node
+// as if it were work to run.
+func releaseCandidate(s *State, owner string) *TopoEntry {
+	var mine []*TopoEntry
+	for _, te := range s.TopologyList() {
+		if te.OwnerID != owner || !te.Active || te.Link.Disabled {
+			continue
+		}
+		mine = append(mine, te)
+	}
+	if len(mine) == 0 {
+		return nil
+	}
+	sort.Slice(mine, func(i, j int) bool {
+		if mine[i].Link.RemotePort != mine[j].Link.RemotePort {
+			return mine[i].Link.RemotePort < mine[j].Link.RemotePort
+		}
+		return mine[i].TaskID < mine[j].TaskID
+	})
+	return mine[0]
+}
+
+// landed reports whether the forward this node released has actually left it.
+//
+// The test is "no longer owned by me", NOT "no entry exists": the ring's normal
+// claim picks the forward up within a round or two, so an existence test would
+// see the new owner's entry and wait forever for a state that has already
+// happened.
+func (b *Balancer) landed() bool {
+	m := b.releasing
+	if m == nil {
+		return true
+	}
+	owner := b.engine.state.TopologyOwner(&Task{
+		Local:  store.Local{Name: m.local},
+		Remote: store.Remote{Name: m.remote},
+		Link:   store.Link{RemotePort: m.port},
+	})
+	return owner != b.engine.ID
+}
+
+// Step advances the balancer by one cycle and returns a human-readable
+// description when it did (or, in dry-run, would have done) something.
+func (b *Balancer) Step(now int64) string {
+	if b.releasing != nil {
+		if b.landed() {
+			done := b.releasing
+			b.releasing = nil
+			return "gave up " + done.local + ":" + strconv.Itoa(done.port) + " for rebalancing"
+		}
+		if now-b.releasing.startedAt > releaseDeadlineSec {
+			log.Printf("ring[%s] release of %s:%d not confirmed within %ds; clearing (forward is still ours)",
+				b.engine.ID, b.releasing.local, b.releasing.port, releaseDeadlineSec)
+			b.releasing = nil
+		}
+		return ""
+	}
+
+	pick, p, blocked := b.assess(now)
+	if pick == nil {
+		b.reported = ""
+		return ""
+	}
+	desc := pick.Local.Name + ":" + strconv.Itoa(pick.Link.RemotePort) +
+		" (" + strconv.Itoa(p.mine) + " owned > " + strconv.Itoa(p.share) + " share of " +
+		strconv.Itoa(p.total) + " over " + strconv.Itoa(p.alive) + " nodes)"
+
+	switch b.mode {
+	case BalanceOff:
+		return ""
+	case BalanceDryRun:
+		// Log a change of mind, not the same line every tick.
+		if b.reported == desc {
+			return ""
+		}
+		b.reported = desc
+		if blocked != "" {
+			return "would release " + desc + " [blocked: " + blocked + "]"
+		}
+		return "would release " + desc
+	}
+
+	// Mode is ON. Blocked reasons still stop us.
+	if blocked != "" {
+		return ""
+	}
+	t := b.engine.ReleaseTask(pick.Local, pick.Remote, pick.Link)
+	if t == nil {
+		return ""
+	}
+	b.releasing = &releaseState{
+		local: pick.Local.Name, remote: pick.Remote.Name,
+		port: pick.Link.RemotePort, startedAt: now,
+	}
+	b.lastReleaseAt = now
+	b.reported = ""
+	return "releasing " + desc
+}
+
+// Watch drives Step on a ticker until ctx is cancelled.
+//
+// It runs on EVERY node, not just the leader. Balancing here is not a cluster
+// decision being delegated to one node — it is each node applying the same
+// replicated rule to its own forwards. Two nodes never contend for the same
+// forward because a node can only release what it owns.
+//
+// The first check happens after one interval rather than immediately: a node
+// that has just booted adopts the ring's topology and its own load in the first
+// token round, and deciding before that would act on a half-known pool.
+func (b *Balancer) Watch(ctx context.Context, interval time.Duration) {
+	tick := time.NewTicker(interval)
+	defer tick.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-tick.C:
+		}
+		if msg := b.Step(time.Now().Unix()); msg != "" {
+			log.Printf("ring[%s] balance: %s", b.engine.ID, msg)
+		}
+	}
+}
+
+// ---- read-only report for the API/UI ----
+
+// BalanceReport is what this node thinks about balancing right now.
+//
+// It is per-node by nature: every member has its own answer, and the answer that
+// matters is the one on the node doing (or refusing to do) the releasing. The
+// peer list is included because "am I over my share?" is only meaningful next to
+// what everyone else holds.
+type BalanceReport struct {
+	Mode       string `json:"mode"`       // off | dry-run | on
+	ModeSource string `json:"modeSource"` // env | stored
+	Enabled    bool   `json:"enabled"`    // mode != off
+	DryRun     bool   `json:"dryRun"`     // mode != on
+	// CanControl: this node offers the mode switch. False for a read-only
+	// principal, so the server does not advertise a control the UI gates anyway.
+	CanControl bool   `json:"canControl"`
+	IsMember   bool   `json:"isMember"`
+	Self       string `json:"self"`
+
+	SelfForwards  int  `json:"selfForwards"`
+	Share         int  `json:"share"`
+	TotalForwards int  `json:"totalForwards"`
+	AliveNodes    int  `json:"aliveNodes"`
+	SelfOverShare bool `json:"selfOverShare"`
+
+	// WouldRelease is what this node would give up (present whenever it is over
+	// its share, including while blocked — see BlockedBy).
+	WouldRelease *BalanceReleasePlan `json:"wouldRelease,omitempty"`
+	// Releasing is the release this node has published and is awaiting.
+	Releasing *BalanceReleasePlan `json:"releasing,omitempty"`
+	// BlockedBy explains why an over-share node is not acting yet.
+	BlockedBy string `json:"blockedBy,omitempty"`
+
+	CooldownSec   int64 `json:"cooldownSec"`
+	LastReleaseAt int64 `json:"lastReleaseAt,omitempty"`
+
+	// Peers: every member's replicated load, for context.
+	Nodes []BalanceNodeReport `json:"nodes"`
+}
+
+type BalanceNodeReport struct {
+	ID       string  `json:"id"`
+	Alive    bool    `json:"alive"`
+	IsLeader bool    `json:"isLeader"`
+	Forwards int     `json:"forwards"`
+	MemPct   float64 `json:"memPct"`
+	NetPct   float64 `json:"netPct"`
+	Score    float64 `json:"score"`
+}
+
+type BalanceReleasePlan struct {
+	Local      string `json:"local"`
+	Remote     string `json:"remote"`
+	RemotePort int    `json:"remotePort"`
+	Owner      string `json:"owner"`
+}
+
+// Report renders the current view. It never mutates anything — in particular it
+// does NOT touch lastReleaseAt or the release state, so an operator polling this
+// cannot start, delay or cancel a release. (A read-only endpoint that perturbs
+// the behaviour it reports is worse than no endpoint.)
+func (b *Balancer) Report() BalanceReport {
+	now := time.Now().Unix()
+	rep := BalanceReport{
+		Mode:        b.mode,
+		ModeSource:  b.ModeSource(),
+		Enabled:     b.mode != BalanceOff,
+		DryRun:      b.mode != BalanceOn,
+		CanControl:  b.persist != nil,
+		IsMember:    b.engine.IsMember(),
+		Self:        b.engine.ID,
+		CooldownSec: b.cooldownSec,
+	}
+	p := b.pool()
+	rep.SelfForwards, rep.Share = p.mine, p.share
+	rep.TotalForwards, rep.AliveNodes = p.total, p.alive
+	rep.SelfOverShare = p.mine > p.share
+
+	if pick, _, blocked := b.assess(now); pick != nil {
+		rep.WouldRelease = &BalanceReleasePlan{
+			Local: pick.Local.Name, Remote: pick.Remote.Name,
+			RemotePort: pick.Link.RemotePort, Owner: pick.OwnerID,
+		}
+		rep.BlockedBy = blocked
+	}
+	if m := b.releasing; m != nil {
+		rep.Releasing = &BalanceReleasePlan{
+			Local: m.local, Remote: m.remote, RemotePort: m.port, Owner: b.engine.ID,
+		}
+		rep.BlockedBy = "in-flight"
+	}
+	if b.lastReleaseAt > 0 {
+		rep.LastReleaseAt = b.lastReleaseAt
+	}
+	for i := range b.engine.state.Nodes {
+		n := &b.engine.state.Nodes[i]
+		rep.Nodes = append(rep.Nodes, BalanceNodeReport{
+			ID: n.ID, Alive: n.Alive, IsLeader: n.IsLeader,
+			Forwards: n.Load.Forwards, MemPct: n.Load.MemPct, NetPct: n.Load.NetPct,
+			Score: n.Load.Score(),
+		})
+	}
+	return rep
+}

internal/cluster/ring_disabled_test.go

@@ -0,0 +1,195 @@
+package cluster
+
+import (
+	"context"
+	"testing"
+
+	"webui4frpc/internal/store"
+)
+
+// staleTokenWith copies the engine's current state but rewrites the given
+// forward's disabled flag, standing in for a token a neighbour captured before
+// this node's decision.
+func staleTokenWith(eng *Engine, local, remote string, port int, disabled bool) State {
+	s := eng.state
+	out := State{
+		LeaderID: s.LeaderID,
+		Nodes:    s.Nodes,
+		Cycle:    s.Cycle,
+		Topology: map[string]*TopoEntry{},
+	}
+	for id, e := range s.Topology {
+		cp := *e
+		if cp.Local.Name == local && cp.Remote.Name == remote && cp.Link.RemotePort == port {
+			cp.Link.Disabled = disabled
+			cp.Active = !disabled
+		}
+		out.Topology[id] = &cp
+	}
+	return out
+}
+
+// TestLocalDisableSurvivesAdoption is the regression test for the wipe that made
+// the ring's authority unachievable.
+//
+// OnToken/AdoptState replace e.state wholesale. A stop decided between two token
+// cycles was therefore erased by the next adoption whenever the incoming token
+// had been captured before the decision — so the flag never reached the other
+// members and the forward kept running. Written as a probe BEFORE the fix: it
+// reported "after adopting a stale token: known=true disabled=false".
+func TestLocalDisableSurvivesAdoption(t *testing.T) {
+	eng := newTestEngine("n1", true)
+	eng.state.AddPending(store.Local{Name: "web"}, store.Remote{Name: "frps1"}, store.Link{RemotePort: 18081})
+	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 1, State: eng.state, SentAt: 1}); err != nil {
+		t.Fatal(err)
+	}
+
+	// Local decision: stop it.
+	eng.UpdateTopologyDisabled("web", "frps1", 18081, true)
+	if d, _ := eng.TopologyDisabled("web", "frps1", 18081); !d {
+		t.Fatal("precondition: the local stop should be visible")
+	}
+
+	// A token captured BEFORE the decision arrives.
+	stale := staleTokenWith(eng, "web", "frps1", 18081, false)
+	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 2, State: stale, SentAt: 2}); err != nil {
+		t.Fatal(err)
+	}
+	// Assert on the STATE, not the TopologyDisabled() accessor: the accessor
+	// short-circuits to the pending local decision, so it would report the stop
+	// even when the state about to be forwarded still says "enabled". What
+	// matters is that the token this node sends on carries the decision.
+	assertForwardedDisabled(t, eng, "web", "frps1", 18081, true)
+}
+
+// assertForwardedDisabled checks the disabled value a DOWNSTREAM node would see
+// in the token this engine forwards — the value that actually propagates.
+func assertForwardedDisabled(t *testing.T, eng *Engine, local, remote string, port int, want bool) {
+	t.Helper()
+	// Model one hop: a fresh engine adopts exactly what this one would send.
+	peer := newTestEngine("peer", false)
+	peer.AdoptState(eng.state)
+	got, known := peer.TopologyDisabled(local, remote, port)
+	if !known {
+		t.Fatalf("peer learned no entry for %s→%s:%d", local, remote, port)
+	}
+	if got != want {
+		t.Fatalf("the forwarded token carries disabled=%v, want %v — the decision did not "+
+			"propagate, so the other members would keep the old state", got, want)
+	}
+}
+
+// TestLocalReEnableSurvivesAdoption: a re-enable needs the same protection. The
+// value false is easy to overlook — nothing "looks" stopped, yet losing it just
+// as silently strands the forward in the stopped state.
+func TestLocalReEnableSurvivesAdoption(t *testing.T) {
+	eng := newTestEngine("n1", true)
+	eng.state.AddPending(store.Local{Name: "web"}, store.Remote{Name: "frps1"}, store.Link{RemotePort: 18081})
+	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 1, State: eng.state, SentAt: 1}); err != nil {
+		t.Fatal(err)
+	}
+	eng.UpdateTopologyDisabled("web", "frps1", 18081, true)
+	eng.UpdateTopologyDisabled("web", "frps1", 18081, false)
+
+	// A token that still says disabled (captured mid-stop) arrives.
+	mid := staleTokenWith(eng, "web", "frps1", 18081, true)
+	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 2, State: mid, SentAt: 2}); err != nil {
+		t.Fatal(err)
+	}
+	assertForwardedDisabled(t, eng, "web", "frps1", 18081, false)
+}
+
+// TestConfirmedDecisionStopsBeingAsserted: once the ring reports the same value
+// the decision is confirmed everywhere, so this node must stop overriding —
+// otherwise it would fight a later decision made elsewhere.
+func TestConfirmedDecisionStopsBeingAsserted(t *testing.T) {
+	eng := newTestEngine("n1", true)
+	eng.state.AddPending(store.Local{Name: "web"}, store.Remote{Name: "frps1"}, store.Link{RemotePort: 18081})
+	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 1, State: eng.state, SentAt: 1}); err != nil {
+		t.Fatal(err)
+	}
+
+	eng.UpdateTopologyDisabled("web", "frps1", 18081, true)
+	if len(eng.localDisabled) != 1 {
+		t.Fatalf("a pending decision should be tracked, got %v", eng.localDisabled)
+	}
+	// The ring comes back agreeing.
+	agreed := staleTokenWith(eng, "web", "frps1", 18081, true)
+	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 2, State: agreed, SentAt: 2}); err != nil {
+		t.Fatal(err)
+	}
+	if len(eng.localDisabled) != 0 {
+		t.Fatalf("a confirmed decision must stop being asserted, still tracking %v", eng.localDisabled)
+	}
+
+	// A peer now decides the opposite; this node must follow the ring.
+	peerSaysEnabled := staleTokenWith(eng, "web", "frps1", 18081, false)
+	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 3, State: peerSaysEnabled, SentAt: 3}); err != nil {
+		t.Fatal(err)
+	}
+	if d, _ := eng.TopologyDisabled("web", "frps1", 18081); d {
+		t.Fatal("this node kept overriding a peer's later decision; the ring is not authoritative")
+	}
+}
+
+// TestLocalDecisionsDoNotAccumulate: the tracking map is swept against the live
+// topology, so a long-lived cluster cannot grow it without bound.
+func TestLocalDecisionsDoNotAccumulate(t *testing.T) {
+	eng := newTestEngine("n1", true)
+	eng.state.AddPending(store.Local{Name: "web"}, store.Remote{Name: "frps1"}, store.Link{RemotePort: 18081})
+	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 1, State: eng.state, SentAt: 1}); err != nil {
+		t.Fatal(err)
+	}
+
+	// Decisions for forwards this node has never seen.
+	eng.UpdateTopologyDisabled("ghost1", "frps1", 1, true)
+	eng.UpdateTopologyDisabled("ghost2", "frps1", 2, true)
+	if len(eng.localDisabled) != 2 {
+		t.Fatalf("expected 2 tracked decisions, got %d", len(eng.localDisabled))
+	}
+	// A cycle runs: neither ghost is in the topology, so both are swept.
+	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 2, State: eng.state, SentAt: 2}); err != nil {
+		t.Fatal(err)
+	}
+	if len(eng.localDisabled) != 0 {
+		t.Fatalf("decisions for forwards outside the topology must be swept, got %v", eng.localDisabled)
+	}
+}
+
+// TestHasSelfInTopology is the wait condition for the startup reconcile: the
+// reconcile rebuilds the workers for forwards THIS node owns, so a node can only
+// act once the adopted state actually names it as an owner.
+//
+// It is deliberately independent of leadership — the bug it guards against was a
+// reconcile that waited for leadership and therefore never ran on followers, even
+// though followers own forwards.
+func TestHasSelfInTopology(t *testing.T) {
+	// Two nodes; the forward is owned by n2, which is NOT the leader. A single-node
+	// ring would make this test pass under either implementation, because the only
+	// member is necessarily the leader — which is exactly how the first version of
+	// this test came to be a false green.
+	eng := newTestEngine("n2", false)
+	eng.state.AddPending(store.Local{Name: "web"}, store.Remote{Name: "frps1"}, store.Link{RemotePort: 18081})
+	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 1, State: eng.state, SentAt: 1}); err != nil {
+		t.Fatal(err)
+	}
+	if eng.IsLeader() {
+		t.Fatal("precondition: n2 must not be the leader in this test")
+	}
+	// n2 claimed it, so it owns a forward — despite not leading.
+	if !eng.HasSelfInTopology() {
+		t.Error("a non-leader that owns a forward reports false — the startup reconcile " +
+			"would skip it and leave its worker down")
+	}
+
+	// A node that owns nothing must report false (so it does not act early on a
+	// stale or empty topology).
+	other := newTestEngine("n9", true)
+	if other.HasSelfInTopology() {
+		t.Error("a node owning nothing must not claim it appears in the topology")
+	}
+	// Guard against a regression where leadership is folded into the predicate.
+	if !eng.HasSelfInTopology() && eng.IsLeader() {
+		t.Error("predicate must not be gated on leadership")
+	}
+}

internal/cluster/ring_engine.go

@@ -7,6 +7,7 @@ import (
 	"encoding/json"
 	"fmt"
 	"log"
+	"strconv"
 	"sync"
 	"time"
 
@@ -18,6 +19,17 @@ import (
 type Handler interface {
 	Claim(ctx context.Context, tk *Task) error
 	Revoke(ctx context.Context, tk *Task) error
+	// Restart re-enables an already-claimed forward and brings its worker back.
+	// Needed because a stop keeps the topology entry, which makes both the
+	// creation path (idempotency guard) and the duplicate-claim guard refuse to
+	// re-create an existing forward.
+	Restart(ctx context.Context, tk *Task) error
+	// Release stops this node's worker for a forward it is giving up for
+	// rebalancing, WITHOUT marking the link stopped. The forward is about to run
+	// on another node, so a stopped flag would be wrong: every node's duplicate
+	// guard reads it and would then refuse to claim the re-published task,
+	// stranding the forward on the node that just gave it up.
+	Release(ctx context.Context, tk *Task) error
 	RuntimeLoad() Load
 }
 
@@ -50,6 +62,24 @@ type Engine struct {
 	// group changes made via HTTP handlers between token cycles are overwritten
 	// by the next state adoption and never propagate to other nodes.
 	topologySync func()
+	// localDisabled records this node's own enabled/disabled decisions that have
+	// not yet been confirmed by the ring. It is what makes the RING authoritative
+	// without losing a decision that was made locally and has not yet had a
+	// chance to reach the other members.
+	//
+	// The problem it solves: OnToken/AdoptState do `e.state = tk.State`, a
+	// wholesale replacement. A stop decided between two token cycles therefore
+	// disappears on the very next adoption if that token was captured before the
+	// decision — the flag never reaches the rest of the ring, and the forward
+	// keeps running. (Verified with a probe before writing this: adopting a
+	// stale token flipped disabled back to false.)
+	//
+	// A key stays in this map until an adoption reports it disabled, at which
+	// point the whole cluster agrees and the key is dropped. Keys for forwards
+	// that vanish from the topology are swept on every update, so the map cannot
+	// grow without bound. The value false is what keeps a RE-ENABLE alive for the
+	// same reason a stop needs protecting.
+	localDisabled map[string]bool
 
 	state State
 	// myAddr maps our Node ID to the address peers dial.
@@ -182,7 +212,9 @@ func (e *Engine) OnToken(ctx context.Context, tk *Token) (*Token, error) {
 	if tk.SentAt > e.lastTokenAt {
 		e.lastTokenAt = tk.SentAt
 	}
-	log.Printf("ring[%s] OnToken cycle=%d", e.ID, tk.Cycle)
+	// Per-round heartbeat — steady state, no information. See debug.go: this
+	// fired ~2x/second and dominated the journal on every node.
+	debugToken("ring[%s] OnToken cycle=%d", e.ID, tk.Cycle)
 
 	// Parallel rhythm timer: operations run while the pace clock ticks.
 	// Delay scales with alive node count (more nodes → lower per-hop delay,
@@ -229,6 +261,11 @@ func (e *Engine) OnToken(ctx context.Context, tk *Token) (*Token, error) {
 	// Re-apply local store overrides (group labels, disabled flags) onto
 	// the freshly adopted topology so they survive state adoption and
 	// propagate to all nodes via the next token forward.
+	// Re-assert decisions this node made locally that the ring has not yet
+	// confirmed, BEFORE the host's topology sync runs: the sync reads
+	// TopologyDisabled to learn the cluster's view, so the local decision must
+	// already be applied or a stop could be reported as "enabled" and dropped.
+	e.reconcileLocalDisabled()
 	if e.topologySync != nil {
 		e.topologySync()
 	}
@@ -375,6 +412,43 @@ func (e *Engine) runCommands(ctx context.Context, tk *Token) error {
 				}
 				continue // owned elsewhere — ride to the owner
 			}
+			if t.Release {
+				// Owner-directed, like a revoke: only the node holding the forward
+				// may give it up. A release executed anywhere else would be an
+				// attempt to move somebody else's load, which cannot work — the
+				// owner's own count is what has to fall, and its topology entry is
+				// local bookkeeping no peer may write.
+				//
+				// A release whose forward is already gone is an idempotent no-op
+				// consumed by the lowest node, so it cannot ride forever if its
+				// owner departed first.
+				if owner := e.state.TopologyOwner(t); owner == e.ID {
+					target = t
+					break
+				} else if owner == "" && selfIsLowest {
+					target = t
+					break
+				}
+				continue // owned elsewhere — ride to the owner
+			}
+			if t.Restart {
+				// Owner-directed, exactly like a revoke: only the node holding the
+				// forward may act. Selecting it anywhere else would let a non-owner
+				// spawn a worker for somebody else's forward (observed live: a
+				// non-owner logged "restarted t4" and ran the forward itself).
+				//
+				// A restart whose forward is already gone is a no-op consumed by the
+				// lowest node, so a task can never ride forever if its owner
+				// departed before seeing it.
+				if owner := e.state.TopologyOwner(t); owner == e.ID {
+					target = t
+					break
+				} else if owner == "" && selfIsLowest {
+					target = t // forward gone — consume and re-enable nothing
+					break
+				}
+				continue // owned elsewhere — ride to the owner
+			}
 			if selfIsLowest {
 				target = t // generic forward create — lowest-load claim
 				break
@@ -388,18 +462,37 @@ func (e *Engine) runCommands(ctx context.Context, tk *Token) error {
 			break
 		}
 		if claimed.Revoke {
-			if e.state.RemoveTopology(claimed.Local.Name, claimed.Remote.Name, claimed.Link.RemotePort) {
-				if e.Handler != nil {
-					if err := e.Handler.Revoke(ctx, claimed); err != nil {
-						log.Printf("ring[%s] revoke %s: %v", e.ID, claimed.ID, err)
-					}
-				}
-				if e.Log != nil {
-					_, _ = e.Log.Append(e.ID, LogForwardRemove, map[string]any{
-						"taskId": claimed.ID, "local": claimed.Local.Name, "remote": claimed.Remote.Name,
-					})
+			// A user-requested stop is recorded as a FLAG on the topology entry,
+			// not as a removal. The entry is the carrier that makes the decision
+			// travel: TopoEntry.Link is a full store.Link and State rides the
+			// token every cycle, so keeping the entry is what lets a stop reach
+			// every member — including a node that was offline when the stop was
+			// issued. Removing the entry instead left the flag nowhere to live,
+			// which is why a stop used to converge only if the owner happened to
+			// be online to receive a one-shot revoke task.
+			//
+			// Marking Active=false is also what makes the entry inert for the
+			// rebalancing paths: OfflineReassign() skips inactive entries, and
+			// the startup reconcile skips them too, so a stopped forward is
+			// neither re-queued on a node departure nor re-claimed on a restart.
+			//
+			// Handler.Revoke still runs unconditionally — it is what actually
+			// stops the per-forward worker, and it must run even when there was no
+			// entry to mark (gating it on RemoveTopology()'s result made stops of
+			// already-absent forwards silently do nothing, so their workers ran
+			// forever: observed live as thousands of connection-refused lines
+			// against a local service that was intentionally down).
+			marked := e.state.UpdateTopologyDisabled(claimed.Local.Name, claimed.Remote.Name, claimed.Link.RemotePort, true)
+			if e.Handler != nil {
+				if err := e.Handler.Revoke(ctx, claimed); err != nil {
+					log.Printf("ring[%s] revoke %s: %v", e.ID, claimed.ID, err)
 				}
 			}
+			if marked && e.Log != nil {
+				_, _ = e.Log.Append(e.ID, LogForwardStop, map[string]any{
+					"taskId": claimed.ID, "local": claimed.Local.Name, "remote": claimed.Remote.Name,
+				})
+			}
 			continue
 		}
 		if claimed.RemoveNode != "" && claimed.RemoveNode == e.ID {
@@ -465,10 +558,90 @@ func (e *Engine) runCommands(ctx context.Context, tk *Token) error {
 		// different node. Safe for OfflineReassign: that path deletes the
 		// topology entry BEFORE re-queueing, so TopologyOwner returns "" and
 		// the legitimate re-claim passes through.
-		if owner := e.state.TopologyOwner(claimed); owner != "" {
-			log.Printf("ring[%s] drop duplicate task %s: %s→%s:%d already owned by %s",
-				e.ID, claimed.ID, claimed.Local.Name, claimed.Remote.Name,
-				claimed.Link.RemotePort, owner)
+		//
+		// A RESTART task is the deliberate exception this guard must not eat: it
+		// exists precisely to re-activate a forward whose entry is still present
+		// (a stop keeps the entry as the flag's carrier), so applying the guard
+		// would swallow every re-enable and leave the forward stopped forever.
+		// A RELEASE is the second deliberate exception, for the same reason as a
+		// restart: it is addressed to the node that ALREADY owns the forward, so
+		// the entry is still there by definition. Applying the guard would drop
+		// every release and a node could never shed anything.
+		if !claimed.Restart && !claimed.Release {
+			if owner := e.state.TopologyOwner(claimed); owner != "" {
+				log.Printf("ring[%s] drop duplicate task %s: %s→%s:%d already owned by %s",
+					e.ID, claimed.ID, claimed.Local.Name, claimed.Remote.Name,
+					claimed.Link.RemotePort, owner)
+				continue
+			}
+		}
+		if claimed.Release {
+			// Only reached when this node owns the forward (or it is already gone
+			// and we are the fallback consumer — see the selection loop).
+			owner := e.state.TopologyOwner(claimed)
+			if owner != e.ID {
+				// The forward vanished before we got here; nothing to give up.
+				continue
+			}
+			// Stop our worker. Deliberately NOT Handler.Revoke: that marks the
+			// link stopped, and the forward is about to run elsewhere — a
+			// "stopped" flag would make every node treat the re-published task as
+			// a duplicate and refuse to claim it.
+			if e.Handler != nil {
+				if err := e.Handler.Release(ctx, claimed); err != nil {
+					log.Printf("ring[%s] release %s: %v", e.ID, claimed.ID, err)
+				}
+			}
+			// Drop the entry so the forward has no owner. This is what makes the
+			// load actually fall (ForwardsOwnedBy counts entries) and what lets a
+			// fresh creation task pass the duplicate guard above.
+			e.RemoveTopologyEntry(claimed.Local.Name, claimed.Remote.Name, claimed.Link.RemotePort)
+			// Hand it back to the ring as an ordinary unclaimed forward. It is
+			// published by THIS node so it rides the very token we are holding —
+			// no coordinator, no addressed hand-off.
+			link := claimed.Link
+			link.Disabled = false
+			if t := e.SubmitTask(claimed.Local, claimed.Remote, link); t != nil {
+				log.Printf("ring[%s] released %s: %s→%s:%d back to the ring as %s",
+					e.ID, claimed.ID, claimed.Local.Name, claimed.Remote.Name,
+					claimed.Link.RemotePort, t.ID)
+			} else {
+				log.Printf("ring[%s] released %s but the re-publish was deduped",
+					e.ID, claimed.ID)
+			}
+			if e.Log != nil {
+				_, _ = e.Log.Append(e.ID, LogForwardStop, map[string]any{
+					"taskId": claimed.ID, "local": claimed.Local.Name,
+					"remote": claimed.Remote.Name, "released": true,
+				})
+			}
+			continue
+		}
+		if claimed.Restart {
+			// Only reached when this node owns the forward (or it is already gone
+			// and we are the fallback consumer — see the selection loop).
+			owner := e.state.TopologyOwner(claimed)
+			e.state.UpdateTopologyDisabled(claimed.Local.Name, claimed.Remote.Name, claimed.Link.RemotePort, false)
+			if owner != e.ID {
+				// Forward vanished before we got here: nothing to re-enable.
+				log.Printf("ring[%s] discard restart %s: no owner", e.ID, claimed.ID)
+				continue
+			}
+			if e.Handler != nil {
+				if err := e.Handler.Restart(ctx, claimed); err != nil {
+					e.state.PendingTasks[claimed.ID] = claimed
+					log.Printf("ring[%s] restart %s failed: %v", e.ID, claimed.ID, err)
+					break
+				}
+			}
+			if e.Log != nil {
+				_, _ = e.Log.Append(e.ID, LogForwardStart, map[string]any{
+					"taskId": claimed.ID, "local": claimed.Local.Name, "remote": claimed.Remote.Name,
+				})
+			}
+			// Logging is the app's job (RestartFn reports the actual worker
+			// outcome); the ring layer stays quiet so one restart is one line
+			// rather than two identical ones.
 			continue
 		}
 		if e.Handler != nil {
@@ -724,6 +897,11 @@ func (e *Engine) AdoptState(s State) {
 	for id, t := range kept {
 		e.state.PendingTasks[id] = t
 	}
+	// Re-assert decisions this node made locally that the ring has not yet
+	// confirmed, BEFORE the host's topology sync runs: the sync reads
+	// TopologyDisabled to learn the cluster's view, so the local decision must
+	// already be applied or a stop could be reported as "enabled" and dropped.
+	e.reconcileLocalDisabled()
 	if e.topologySync != nil {
 		e.topologySync()
 	}
@@ -792,6 +970,14 @@ func (e *Engine) RemoveNode(nodeID string) *Task {
 	return e.state.AddRemoveNode(nodeID)
 }
 
+// RemoveTopologyEntry drops the active topology entry for a forward identified
+// by its natural key. Exposed so the app's claim path can retire an entry it
+// refuses to serve (see ClaimFn: a disabled forward must stop looking active,
+// otherwise the startup reconcile keeps re-claiming it on every restart).
+func (e *Engine) RemoveTopologyEntry(local, remote string, port int) bool {
+	return e.state.RemoveTopology(local, remote, port)
+}
+
 // RevokeTask publishes a revocation for an established forward through the
 // same token channel; the owning node stops the worker and drops topology.
 func (e *Engine) RevokeTask(local store.Local, remote store.Remote, link store.Link) *Task {
@@ -799,14 +985,107 @@ func (e *Engine) RevokeTask(local store.Local, remote store.Remote, link store.L
 }
 
 func (e *Engine) SubmitTask(local store.Local, remote store.Remote, link store.Link) *Task {
-	if e.HasTask(local.Name, remote.Name, link.RemotePort) {
+	// Guard on an ACTIVE task only. A topology entry that is merely marked
+	// disabled is kept around as the carrier for the stop flag, so counting it
+	// here would make this a permanent no-op and a stopped forward could never
+	// be started again.
+	if e.HasActiveTask(local.Name, remote.Name, link.RemotePort) {
 		return nil
 	}
 	return e.state.AddPending(local, remote, link)
 }
 
-// HasTask reports whether a forward with the same local/remote/remotePort is
-// already pending or active in the topology (idempotency guard for resaves).
+// HasTopologyEntry reports whether the forward has a topology entry (whether or
+// not it is marked disabled). A stop keeps the entry as the flag's carrier, so
+// "has an entry" is what distinguishes "already claimed by someone" from "never
+// submitted" — the distinction startForward needs.
+func (e *Engine) HasTopologyEntry(local, remote string, port int) bool {
+	_, found := e.state.TopologyDisabled(local, remote, port)
+	return found
+}
+
+// SubmitRestart publishes a task asking the current owner of an already-claimed
+// forward to bring its worker back up.
+//
+// Neither of the existing channels can express "restart what already exists":
+// SubmitTask is guarded by HasActiveTask and the entry is still present, so it
+// dedupes; and the claim path drops any task for a forward that already has an
+// owner (its defence against duplicate-claim collisions). Re-enabling a stopped
+// forward therefore needs its own task kind, which the owner applies without
+// re-running the claim bookkeeping.
+func (e *Engine) SubmitRestart(local store.Local, remote store.Remote, link store.Link) *Task {
+	// The task must not carry the stop: the whole point is to re-enable.
+	link.Disabled = false
+	t := &Task{
+		ID:      e.state.NextTaskID(),
+		Local:   local,
+		Remote:  remote,
+		Link:    link,
+		Created: time.Now().Unix(),
+		Restart: true,
+	}
+	if e.state.PendingTasks == nil {
+		e.state.PendingTasks = map[string]*Task{}
+	}
+	e.state.PendingTasks[t.ID] = t
+	return t
+}
+
+// ReleaseTask publishes a rebalancing task asking the OWNER of this forward to
+// give it up. The task rides the token like any other command; it is not
+// addressed to a destination, because the ring's normal lowest-load claim
+// decides where the forward goes next.
+//
+// It goes through AddPending rather than writing PendingTasks directly. That is
+// not cosmetic: OnToken rebuilds local-only pending commands from the ids it
+// never published, and only the Add* helpers record them — a task injected by
+// hand was silently dropped at the next state adoption, so it never reached the
+// ring at all.
+func (e *Engine) ReleaseTask(local store.Local, remote store.Remote, link store.Link) *Task {
+	link.Disabled = false
+	t := e.state.AddPending(local, remote, link)
+	if t == nil {
+		return nil
+	}
+	t.Release = true
+	return t
+}
+
+// HasActiveTask reports whether a forward with this key is genuinely in flight
+// or running: a non-revocation pending task, or a topology entry that is not
+// marked disabled.
+//
+// It is deliberately narrower than HasTask. Under mark-don't-remove a stopped
+// forward keeps its topology entry (that entry is what carries the flag between
+// nodes), so "an entry exists" no longer means "this forward is active".
+// SubmitTask must use THIS predicate, otherwise starting a stopped forward —
+// startForward() publishes the enable and then calls SubmitTask — would be
+// swallowed by its own idempotency guard.
+func (e *Engine) HasActiveTask(local, remote string, port int) bool {
+	for _, t := range e.state.PendingList() {
+		if t.Revoke {
+			continue // a revocation is the opposite of an active forward
+		}
+		if t.Local.Name == local && t.Remote.Name == remote && t.Link.RemotePort == port {
+			return true
+		}
+	}
+	for _, t := range e.state.TopologyList() {
+		if t.Link.Disabled {
+			continue // stopped; the entry is only a flag carrier
+		}
+		if t.Local.Name == local && t.Remote.Name == remote && t.Link.RemotePort == port {
+			return true
+		}
+	}
+	return false
+}
+
+// HasTask reports whether ANY record for this forward exists — a pending task
+// (including an in-flight revocation) or a topology entry (including one marked
+// disabled). This is the "have we already handled this key" question, used to
+// avoid re-issuing work on repeated canvas saves. For "is it actually running"
+// use HasActiveTask instead.
 func (e *Engine) HasTask(local, remote string, port int) bool {
 	for _, t := range e.state.PendingList() {
 		if t.Local.Name == local && t.Remote.Name == remote && t.Link.RemotePort == port {
@@ -828,6 +1107,78 @@ func (e *Engine) UpdateTopologyGroup(local, remote string, port int, group strin
 	return e.state.UpdateTopologyGroup(local, remote, port, group)
 }
 
+// UpdateTopologyDisabled marks a forward enabled/disabled in the topology so
+// the decision rides the next token cycle to every member. Paired with
+// TopologyDisabled, which the adoption hook uses to learn peers' decisions.
+//
+// The decision is also remembered locally until the ring confirms it, because
+// adoption replaces the whole state: a token captured before this call would
+// otherwise wash the decision out on the next cycle and it would never reach
+// the other members. See the localDisabled field.
+func (e *Engine) UpdateTopologyDisabled(local, remote string, port int, disabled bool) bool {
+	if e.localDisabled == nil {
+		e.localDisabled = map[string]bool{}
+	}
+	e.localDisabled[disableKey(local, remote, port)] = disabled
+	return e.state.UpdateTopologyDisabled(local, remote, port, disabled)
+}
+
+// TopologyDisabled reports the cluster's view of a forward's disabled flag.
+// The bool is false when the forward has no topology entry (nothing to learn).
+//
+// A pending local decision takes precedence over the adopted state: it has not
+// had a chance to reach the other members yet, and reporting the adopted value
+// would make the local store (and the status page) disagree with the user's
+// most recent action.
+func (e *Engine) TopologyDisabled(local, remote string, port int) (bool, bool) {
+	if d, pending := e.localDisabled[disableKey(local, remote, port)]; pending {
+		return d, true
+	}
+	return e.state.TopologyDisabled(local, remote, port)
+}
+
+// reconcileLocalDisabled re-asserts this node's not-yet-confirmed disabled
+// decisions onto the freshly adopted state. Called right after e.state is
+// replaced, and paired with reconcileTopologySync (which pushes those decisions
+// out to the ring).
+//
+// A decision is dropped once the adopted state reports the SAME value: at that
+// point every member agrees and there is nothing left to protect. Entries whose
+// forward no longer exists in the topology are dropped too, so the map tracks
+// only live disagreements.
+func (e *Engine) reconcileLocalDisabled() {
+	if len(e.localDisabled) == 0 {
+		return
+	}
+	for id, te := range e.state.Topology {
+		_ = id
+		k := disableKey(te.Local.Name, te.Remote.Name, te.Link.RemotePort)
+		want, pending := e.localDisabled[k]
+		if !pending {
+			continue
+		}
+		if te.Link.Disabled == want {
+			// The ring caught up (or agreed independently): stop tracking.
+			delete(e.localDisabled, k)
+			continue
+		}
+		// Still divergent: keep asserting our decision onto the adopted state.
+		te.Link.Disabled = want
+		te.Active = !want
+	}
+	// Forget decisions for forwards that left the topology entirely — otherwise
+	// a long-lived cluster would accumulate dead keys.
+	live := make(map[string]struct{}, len(e.state.Topology))
+	for _, te := range e.state.Topology {
+		live[disableKey(te.Local.Name, te.Remote.Name, te.Link.RemotePort)] = struct{}{}
+	}
+	for k := range e.localDisabled {
+		if _, ok := live[k]; !ok {
+			delete(e.localDisabled, k)
+		}
+	}
+}
+
 // IsLeader reports whether this node is the current ring leader.
 func (e *Engine) IsLeader() bool { return e.state.LeaderID == e.ID }
 
@@ -850,6 +1201,14 @@ func (e *Engine) SetPeerPersist(fn func(peersJSON string) error) { e.peerPersist
 // HTTP handlers are overwritten by the next e.state = tk.State.
 func (e *Engine) SetTopologySync(fn func()) { e.topologySync = fn }
 
+// disableKey builds the map key for a forward's disabled decision: the natural
+// triple (local, remote, remotePort), which is how every other part of the code
+// identifies a forward. A NUL separator keeps it unambiguous for names that
+// could otherwise collide across the boundaries.
+func disableKey(local, remote string, port int) string {
+	return local + "\x00" + remote + "\x00" + strconv.Itoa(port)
+}
+
 // persistPeers extracts all alive peers (addr + nodeKey, excluding self)
 // from the current ring state and persists them via the peerPersist callback.
 // Called on every token cycle (OnToken) and on AdoptState so a crashed node
@@ -938,3 +1297,20 @@ func (e *Engine) LeaderAddr() string {
 	}
 	return e.state.Nodes[i].Addr
 }
+
+// HasSelfInTopology reports whether the current topology names this node as the
+// owner of at least one active forward.
+//
+// The startup reconcile uses it as its wait condition: a node can only rebuild
+// the workers for forwards it owns once the adopted state says it owns them.
+// Leadership is deliberately NOT part of the test — followers own forwards too,
+// and gating the reconcile on leadership left them permanently unable to restore
+// their own workers.
+func (e *Engine) HasSelfInTopology() bool {
+	for _, te := range e.state.TopologyList() {
+		if te.OwnerID == e.ID && te.Active {
+			return true
+		}
+	}
+	return false
+}

internal/cluster/ring_engine_test.go

@@ -8,9 +8,18 @@ import (
 )
 
 type fakeHandler struct {
-	load   Load
-	claim  func(ctx context.Context, tk *Task) error
-	revoke func(ctx context.Context, tk *Task) error
+	load    Load
+	claim   func(ctx context.Context, tk *Task) error
+	revoke  func(ctx context.Context, tk *Task) error
+	restart func(ctx context.Context, tk *Task) error
+	release func(ctx context.Context, tk *Task) error
+}
+
+func (h *fakeHandler) Release(ctx context.Context, tk *Task) error {
+	if h.release != nil {
+		return h.release(ctx, tk)
+	}
+	return nil
 }
 
 func (h *fakeHandler) Revoke(ctx context.Context, tk *Task) error {
@@ -20,6 +29,13 @@ func (h *fakeHandler) Revoke(ctx context.Context, tk *Task) error {
 	return nil
 }
 
+func (h *fakeHandler) Restart(ctx context.Context, tk *Task) error {
+	if h.restart != nil {
+		return h.restart(ctx, tk)
+	}
+	return nil
+}
+
 func (h *fakeHandler) RuntimeLoad() Load { return h.load }
 func (h *fakeHandler) Claim(ctx context.Context, tk *Task) error {
 	if h.claim != nil {

internal/cluster/ring_handler.go

@@ -17,6 +17,16 @@ type AppHandler struct {
 	ClaimFn func(ctx context.Context, tk *Task) error
 	// RevokeFn cancels the forward on this node (stop worker + drop from store).
 	RevokeFn func(ctx context.Context, tk *Task) error
+	// RestartFn re-enables an already-claimed forward on this node (clear the
+	// stopped flag + bring the worker back). Required because a stop keeps the
+	// topology entry, so the normal claim path will not re-create it.
+	RestartFn func(ctx context.Context, tk *Task) error
+	// ReleaseFn stops this node's worker for a forward being handed back to the
+	// ring, leaving the link ENABLED. When nil, Release falls back to RevokeFn —
+	// which also stops the worker, but additionally marks the link stopped; that
+	// is the pre-existing stop semantics and is wrong for a rebalance, so the app
+	// should provide this.
+	ReleaseFn func(ctx context.Context, tk *Task) error
 }
 
 // RuntimeLoad implements Handler.
@@ -44,6 +54,30 @@ func (a *AppHandler) Revoke(ctx context.Context, tk *Task) error {
 	return a.RevokeFn(ctx, tk)
 }
 
+// Release implements Handler.
+func (a *AppHandler) Release(ctx context.Context, tk *Task) error {
+	if a.ReleaseFn != nil {
+		return a.ReleaseFn(ctx, tk)
+	}
+	// Fall back to the stop path so a host that has not provided ReleaseFn still
+	// stops the worker (the forward keeps running otherwise). The link ends up
+	// flagged stopped, which the claim path may then refuse — the host is warned
+	// by the ring's own "re-publish was deduped" log line rather than silently
+	// leaving two workers up.
+	if a.RevokeFn != nil {
+		return a.RevokeFn(ctx, tk)
+	}
+	return nil
+}
+
+// Restart implements Handler.
+func (a *AppHandler) Restart(ctx context.Context, tk *Task) error {
+	if a.RestartFn == nil {
+		return nil
+	}
+	return a.RestartFn(ctx, tk)
+}
+
 // SampleMemLoad returns a cheap memory-usage percentage (0..100).
 func SampleMemLoad() float64 {
 	var m runtime.MemStats

internal/cluster/ring_leader.go

@@ -128,7 +128,7 @@ func (e *Engine) forwardToNext(ctx context.Context, tk *Token) error {
 		if e.send == nil {
 			return nil
 		}
-		log.Printf("ring[%s] forward cycle=%d to %s", e.ID, tk.Cycle, next)
+		debugToken("ring[%s] forward cycle=%d to %s", e.ID, tk.Cycle, next)
 		err := e.send(ctx, next, tk)
 		if err == nil {
 			if e.state.LeaderID == e.ID {
@@ -146,7 +146,14 @@ func (e *Engine) forwardToNext(ctx context.Context, tk *Token) error {
 		// the token is lost — start fresh as the new leader (plan §leader
 		// 补充/监控: "上家邻居探测到 leader 崩溃 → 自身成为新 leader").
 		if next == e.state.LeaderID {
-			e.becomeLeader()
+			if !e.becomeLeader() {
+				// Another live leader exists (we are the partitioned side):
+				// keep its ring intact and let it re-drive the token. Re-raise
+				// the node we just marked dead only if the send really was the
+				// problem — otherwise we would paper over a real outage.
+				log.Printf("ring[%s] not promoting: deferring to live leader", e.ID)
+				return nil
+			}
 			if e.Log != nil {
 				_, _ = e.Log.Append(e.ID, LogLeaderChange, map[string]string{"leader": e.ID})
 			}
@@ -161,13 +168,42 @@ func (e *Engine) forwardToNext(ctx context.Context, tk *Token) error {
 	return nil
 }
 
-// becomeLeader promotes this node (used when the monitored leader dies).
-func (e *Engine) becomeLeader() {
+// becomeLeader promotes this node, but ONLY when the ring agrees there is no
+// live leader to defer to.
+//
+// The guard matters because leadership was previously self-declared: any node
+// that failed to reach the leader it believed in would immediately promote
+// itself. Two nodes partitioned from each other both hit that path and both
+// ended up believing they were the leader — a split brain with two token
+// rings, duplicate claims, and each side reassigning the other's forwards. That
+// was observed live: after a rolling restart, two nodes reported the same
+// leader while marking each other dead, and the topology was emptied into
+// pending on both sides.
+//
+// Before promoting, ask the ring: is any OTHER node currently alive and claiming
+// leadership? If so, we are the partitioned one — demote ourselves and let the
+// live leader keep the ring. A genuine single-leader loss (the real leader is
+// dead or unreachable) leaves no such node, so promotion still happens.
+//
+// CreateCluster does not come through here: it assigns LeaderID directly, so
+// the "no other leader exists" case never needs an override.
+func (e *Engine) becomeLeader() bool {
+	for i := range e.state.Nodes {
+		n := &e.state.Nodes[i]
+		if n.ID == e.ID {
+			continue
+		}
+		if n.Alive && n.IsLeader {
+			log.Printf("ring[%s] decline promotion: %s is alive and is the leader", e.ID, n.ID)
+			return false
+		}
+	}
 	e.state.LeaderID = e.ID
 	for i := range e.state.Nodes {
 		e.state.Nodes[i].IsLeader = e.state.Nodes[i].ID == e.ID
 	}
 	log.Printf("ring[%s] promoted to leader", e.ID)
+	return true
 }
 
 // WatchLeader runs the FALLBACK leader liveness monitor. The PRIMARY path is
@@ -209,20 +245,43 @@ func (e *Engine) WatchLeader(ctx context.Context) {
 				cancel()
 				if err != nil {
 					log.Printf("ring[%s] heartbeat to leader %s failed: %v", e.ID, e.state.LeaderID, err)
-					e.state.MarkOffline(e.state.LeaderID)
-					e.state.OfflineReassign(e.state.LeaderID)
-					e.becomeLeader()
-					// Kick off a fresh cycle: the ring died with the old
-					// leader (no token inflight → WatchTokenLoss won't
-					// fire). Without this the newly promoted leader would
-					// sit idle and the ring would stay dead.
-					e.StartRing(ctx)
+					e.takeOverFromDeadLeader(ctx)
 				}
 			}
 		}
 	}
 }
 
+// takeOverFromDeadLeader handles "the leader I could not reach is gone": mark it
+// dead and, if the ring agrees there is no live leader left, promote this node
+// and re-seed the cycle.
+//
+// The promotion is deliberately conditional, and so is the reassignment. Doing
+// the reassignment unconditionally is what destroyed the live ring: a node on
+// the partitioned side would mark the (still very much alive) leader dead and
+// move ALL of its forwards into pending, so both sides ended up with an empty
+// topology and a queue nobody was serving. Sequencing matters — the forward
+// re-hang is only correct once we have actually taken leadership.
+//
+// Extracted from WatchLeader's loop so the decision is testable (the loop itself
+// is an endless goroutine).
+func (e *Engine) takeOverFromDeadLeader(ctx context.Context) {
+	old := e.state.LeaderID
+	e.state.MarkOffline(old)
+	if !e.becomeLeader() {
+		log.Printf("ring[%s] leader %s unreachable but a live leader remains; "+
+			"declining takeover and leaving its forwards untouched", e.ID, old)
+		return
+	}
+	e.state.OfflineReassign(old)
+	if e.Log != nil {
+		_, _ = e.Log.Append(e.ID, LogLeaderChange, map[string]string{"leader": e.ID})
+	}
+	// The ring died with the old leader (no token inflight → WatchTokenLoss
+	// won't fire). Without this the newly promoted leader would sit idle.
+	e.StartRing(ctx)
+}
+
 // WatchTokenLoss runs the leader's token-loss judge: if a token was sent and
 // does not return within LossTimeout, the leader issues a fresh token (all
 // nodes drop older stamps via the SentAt guard, so at most one circulates).

internal/cluster/ring_log.go

@@ -20,6 +20,8 @@ import (
 // LogKind enumerates operation-log entry types.
 const (
 	LogForwardAdd    = "forward.add"
+	LogForwardStop   = "forward.stop"
+	LogForwardStart  = "forward.start"
 	LogForwardRemove = "forward.remove"
 	LogNodeJoin      = "node.join"
 	LogNodeLeave     = "node.leave"
@@ -140,7 +142,7 @@ func DetailOf(e LogEntry) string {
 	}
 	str := func(key string) string { s, _ := d[key].(string); return s }
 	switch e.Kind {
-	case LogForwardAdd, LogForwardRemove:
+	case LogForwardAdd, LogForwardStop, LogForwardStart, LogForwardRemove:
 		s := str("local") + " → " + str("remote")
 		if id := str("taskId"); id != "" {
 			if len(id) > 8 {

internal/cluster/ring_revoke_test.go

@@ -2,14 +2,22 @@ package cluster
 
 import (
 	"context"
+	"encoding/json"
 	"testing"
 
 	"webui4frpc/internal/store"
 )
 
-// TestRevokeTaskRemovesTopology: a revoke task published to the ring removes
-// the forward from topology, calls Handler.Revoke, and logs forward.remove.
-func TestRevokeTaskRemovesTopology(t *testing.T) {
+// TestRevokeTaskMarksTopologyDisabled: a stop delivered through the ring marks
+// the topology entry disabled instead of deleting it, calls Handler.Revoke, and
+// logs forward.stop.
+//
+// The entry is deliberately KEPT: TopoEntry.Link is the carrier that makes the
+// stop travel, and State rides the token every cycle. Keeping it is what lets a
+// stop reach a member that was offline when the stop was issued — deleting the
+// entry left the flag nowhere to live, so a stop converged only if the owning
+// node happened to be online for a one-shot revoke task.
+func TestRevokeTaskMarksTopologyDisabled(t *testing.T) {
 	revoked := false
 	eng := NewEngine("n1", "n1:7500", "u", "p", "0.71.0", nil,
 		&fakeHandler{load: Load{MemPct: 5, NetPct: 5},
@@ -26,33 +34,92 @@ func TestRevokeTaskRemovesTopology(t *testing.T) {
 	if len(eng.state.TopologyList()) != 1 {
 		t.Fatalf("topology after claim = %+v", eng.state.TopologyList())
 	}
+	if d, _ := eng.state.TopologyDisabled("web", "frps1", 18081); d {
+		t.Fatal("a freshly claimed forward must not start out disabled")
+	}
 
 	// publish a revoke task pointing at the same forward
 	eng.state.AddRevoke(store.Local{Name: "web"}, store.Remote{Name: "frps1"}, store.Link{RemotePort: 18081})
 	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 2, State: eng.state}); err != nil {
 		t.Fatal(err)
 	}
-	if len(eng.state.TopologyList()) != 0 {
-		t.Fatalf("topology after revoke = %+v", eng.state.TopologyList())
+
+	// The entry must SURVIVE, marked disabled, so the flag keeps travelling.
+	if len(eng.state.TopologyList()) != 1 {
+		t.Fatalf("topology entry was removed; the disabled flag would have no carrier: %+v",
+			eng.state.TopologyList())
+	}
+	disabled, known := eng.state.TopologyDisabled("web", "frps1", 18081)
+	if !known {
+		t.Fatal("topology entry missing after revoke")
+	}
+	if !disabled {
+		t.Fatal("revoke did not mark the topology entry disabled")
+	}
+	// Active must follow the flag, since OfflineReassign only re-queues Active
+	// entries — a stopped forward must not be resurrected by a node departure.
+	if eng.state.TopologyList()[0].Active {
+		t.Fatal("a stopped forward must not remain Active (OfflineReassign would re-queue it)")
 	}
 	if !revoked {
 		t.Fatal("Handler.Revoke was not called")
 	}
-	// log should contain forward.remove
-	var sawRemove bool
+	// log should contain forward.stop (not forward.remove — nothing was removed)
+	var sawStop bool
 	for _, e := range eng.Log.Snapshot() {
-		if e.Kind == LogForwardRemove {
-			sawRemove = true
+		if e.Kind == LogForwardStop {
+			sawStop = true
 		}
 	}
-	if !sawRemove {
-		t.Fatalf("log missing forward.remove: %+v", eng.Log.Snapshot())
+	if !sawStop {
+		t.Fatalf("log missing forward.stop: %+v", eng.Log.Snapshot())
+	}
+}
+
+// TestStoppedTopologyEntrySurvivesAdoption is the property that makes the whole
+// mark-don't-remove design work: because the entry persists, a node that adopts
+// ring state learns the stop. This is the offline-member case a one-shot revoke
+// task could never cover.
+func TestStoppedTopologyEntrySurvivesAdoption(t *testing.T) {
+	src := newTestEngine("n1", true)
+	src.state.AddPending(store.Local{Name: "web"}, store.Remote{Name: "frps1"}, store.Link{RemotePort: 18081})
+	if _, err := src.OnToken(context.Background(), &Token{Cycle: 1, State: src.state}); err != nil {
+		t.Fatal(err)
+	}
+	src.state.AddRevoke(store.Local{Name: "web"}, store.Remote{Name: "frps1"}, store.Link{RemotePort: 18081})
+	if _, err := src.OnToken(context.Background(), &Token{Cycle: 2, State: src.state}); err != nil {
+		t.Fatal(err)
+	}
+
+	// A different node adopts the ring state (what e.state = tk.State does).
+	peer := newTestEngine("n2", false)
+	peer.AdoptState(src.state)
+
+	disabled, known := peer.TopologyDisabled("web", "frps1", 18081)
+	if !known {
+		t.Fatal("peer did not receive the topology entry for the stopped forward")
+	}
+	if !disabled {
+		t.Fatal("peer adopted the entry but no longer sees it as disabled — the stop did not propagate")
 	}
 }
 
 // TestRevokeIdempotent: revoking an already-missing forward does not error.
+//
+// It must ALSO still call Handler.Revoke. The handler is what actually stops
+// the per-forward frpc worker; the topology entry is only bookkeeping. Gating
+// the handler on RemoveTopology()'s return value (as this test used to permit)
+// turned every revoke of an already-absent forward into a silent no-op — the
+// worker kept running, which is how a forward the user had stopped kept
+// dialling a local service that was intentionally down, forever.
 func TestRevokeIdempotent(t *testing.T) {
-	eng := newTestEngine("n1", true)
+	revoked := 0
+	eng := NewEngine("n1", "n1:7500", "u", "p", "0.71.0", nil,
+		&fakeHandler{load: Load{MemPct: 5, NetPct: 5},
+			revoke: func(ctx context.Context, tk *Task) error { revoked++; return nil }},
+		func(ctx context.Context, next string, tk *Token) error { return nil },
+		"n1:7500", true, "")
+
 	eng.state.AddRevoke(store.Local{Name: "ghost"}, store.Remote{Name: "frps1"}, store.Link{RemotePort: 1})
 	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 1, State: eng.state}); err != nil {
 		t.Fatalf("revoke missing: %v", err)
@@ -61,4 +128,318 @@ func TestRevokeIdempotent(t *testing.T) {
 	if len(eng.state.TopologyList()) != 0 {
 		t.Fatal("should be empty")
 	}
+	// The stop side-effect must have happened even though there was no entry.
+	if revoked != 1 {
+		t.Fatalf("Handler.Revoke called %d times, want 1 — a revoke with no topology entry "+
+			"must still stop the worker, otherwise stopped forwards keep running", revoked)
+	}
+}
+
+// TestStoppedForwardNotRequeuedOnNodeDeparture pins why marking (rather than
+// deleting) is safe: OfflineReassign only re-queues ACTIVE entries, so a
+// stopped forward is not silently handed to another node when its owner goes
+// away. Deleting the entry, or leaving it Active, would both resurrect it.
+func TestStoppedForwardNotRequeuedOnNodeDeparture(t *testing.T) {
+	eng := newTestEngine("n1", true)
+	eng.state.AddPending(store.Local{Name: "web"}, store.Remote{Name: "frps1"}, store.Link{RemotePort: 18081})
+	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 1, State: eng.state}); err != nil {
+		t.Fatal(err)
+	}
+	eng.state.AddRevoke(store.Local{Name: "web"}, store.Remote{Name: "frps1"}, store.Link{RemotePort: 18081})
+	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 2, State: eng.state}); err != nil {
+		t.Fatal(err)
+	}
+
+	owner := eng.state.TopologyList()[0].OwnerID
+	eng.state.OfflineReassign(owner)
+
+	if n := len(eng.state.PendingList()); n != 0 {
+		t.Fatalf("a stopped forward was re-queued on the owner's departure (pending=%d): %+v",
+			n, eng.state.PendingList())
+	}
+}
+
+// TestSubmitTaskNotBlockedByStoppedEntry is the regression test for the bug the
+// mark-don't-remove change introduced: SubmitTask's idempotency guard used
+// HasTask, which matches a disabled topology entry. Since a stop now KEEPS the
+// entry, any submission for that forward while it is still stopped is swallowed
+// by its own guard — so a stopped forward can never be started again.
+//
+// The probe deliberately submits WITHOUT re-enabling first. Re-enabling would
+// clear the flag and make HasTask and HasActiveTask agree, hiding the defect;
+// the guard has to be exercised at the moment the entry is still stopped.
+func TestSubmitTaskNotBlockedByStoppedEntry(t *testing.T) {
+	eng := newTestEngine("n1", true)
+	eng.state.AddPending(store.Local{Name: "web"}, store.Remote{Name: "frps1"}, store.Link{RemotePort: 18081})
+	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 1, State: eng.state}); err != nil {
+		t.Fatal(err)
+	}
+
+	// stop it -> entry is kept, marked disabled
+	eng.state.AddRevoke(store.Local{Name: "web"}, store.Remote{Name: "frps1"}, store.Link{RemotePort: 18081})
+	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 2, State: eng.state}); err != nil {
+		t.Fatal(err)
+	}
+	if d, _ := eng.state.TopologyDisabled("web", "frps1", 18081); !d {
+		t.Fatal("precondition: forward should be disabled")
+	}
+	// The stopped entry IS findable by the broad predicate — which is exactly why
+	// the narrow one has to exist.
+	if !eng.HasTask("web", "frps1", 18081) {
+		t.Fatal("precondition: the stopped entry should still be findable by HasTask")
+	}
+	if eng.HasActiveTask("web", "frps1", 18081) {
+		t.Fatal("precondition: a stopped forward must NOT count as active")
+	}
+
+	// The regression: a submission must be published for a stopped forward.
+	// With a HasTask guard this returns nil and the forward is stuck forever.
+	tk := eng.SubmitTask(store.Local{Name: "web"}, store.Remote{Name: "frps1"},
+		store.Link{RemotePort: 18081})
+	if tk == nil {
+		t.Fatal("SubmitTask was swallowed by the stopped topology entry — " +
+			"a stopped forward could never be started again")
+	}
+	if tk.Link.RemotePort != 18081 {
+		t.Fatalf("unexpected task: %+v", tk)
+	}
+}
+
+// TestSubmitTaskStillDedupesActiveForward is the other half of the guard: the
+// narrower predicate must not turn SubmitTask into a duplicate-task generator.
+func TestSubmitTaskStillDedupesActiveForward(t *testing.T) {
+	eng := newTestEngine("n1", true)
+	eng.state.AddPending(store.Local{Name: "web"}, store.Remote{Name: "frps1"}, store.Link{RemotePort: 18081})
+	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 1, State: eng.state}); err != nil {
+		t.Fatal(err)
+	}
+	if !eng.HasActiveTask("web", "frps1", 18081) {
+		t.Fatal("precondition: a claimed forward must count as active")
+	}
+	if tk := eng.SubmitTask(store.Local{Name: "web"}, store.Remote{Name: "frps1"},
+		store.Link{RemotePort: 18081}); tk != nil {
+		t.Fatalf("SubmitTask must stay idempotent for an active forward, got %+v", tk)
+	}
+}
+
+// TestStoppedForwardNotRequeuedOnNodeDeparture pins why marking (rather than
+// deleting) is safe: OfflineReassign only re-queues ACTIVE entries, so a
+// stopped forward is not silently handed to another node when its owner goes
+
+// TestAddTopologyRespectsDisabledFlag: a claim that carries a stopped forward
+// must not mark the new entry Active, or it would come back to life.
+func TestAddTopologyRespectsDisabledFlag(t *testing.T) {
+	s := &State{Topology: map[string]*TopoEntry{}}
+	e := s.AddTopology(&Task{
+		ID: "t1", Local: store.Local{Name: "web"}, Remote: store.Remote{Name: "frps1"},
+		Link: store.Link{RemotePort: 18081, Disabled: true},
+	}, "n1")
+	if e.Active {
+		t.Fatal("AddTopology forced Active=true for a disabled claim — the forward would resurrect")
+	}
+	if !e.Link.Disabled {
+		t.Fatal("AddTopology dropped the disabled flag from the link")
+	}
+}
+
+// TestRestartTaskBypassesDuplicateClaimGuard is the engine half of the
+// re-enable path: a restart task for a forward that still HAS a topology entry
+// must reach Handler.Restart.
+//
+// This is precisely what the duplicate-claim guard refuses — it exists to stop
+// a stale task from spawning a second worker for an owned forward, and a restart
+// is indistinguishable from that unless it is an explicit task kind. Marking
+// the entry stopped (instead of deleting it) is what made this necessary: the
+// entry the guard keys on now outlives a stop.
+func TestRestartTaskBypassesDuplicateClaimGuard(t *testing.T) {
+	restarts := 0
+	claims := 0
+	eng := NewEngine("n1", "n1:7500", "u", "p", "0.71.0", nil,
+		&fakeHandler{load: Load{MemPct: 5, NetPct: 5},
+			claim:   func(ctx context.Context, tk *Task) error { claims++; return nil },
+			restart: func(ctx context.Context, tk *Task) error { restarts++; return nil }},
+		func(ctx context.Context, next string, tk *Token) error { return nil },
+		"n1:7500", true, "")
+
+	// Establish + stop: entry survives, marked disabled.
+	eng.state.AddPending(store.Local{Name: "web"}, store.Remote{Name: "frps1"}, store.Link{RemotePort: 18081})
+	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 1, State: eng.state}); err != nil {
+		t.Fatal(err)
+	}
+	eng.state.AddRevoke(store.Local{Name: "web"}, store.Remote{Name: "frps1"}, store.Link{RemotePort: 18081})
+	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 2, State: eng.state}); err != nil {
+		t.Fatal(err)
+	}
+	if d, _ := eng.state.TopologyDisabled("web", "frps1", 18081); !d {
+		t.Fatal("precondition: should be disabled")
+	}
+	entryCount := len(eng.state.TopologyList())
+
+	// Start: publish a restart and run a cycle.
+	eng.state.UpdateTopologyDisabled("web", "frps1", 18081, false)
+	eng.SubmitRestart(store.Local{Name: "web"}, store.Remote{Name: "frps1"},
+		store.Link{RemotePort: 18081})
+	if _, err := eng.OnToken(context.Background(), &Token{Cycle: 3, State: eng.state}); err != nil {
+		t.Fatal(err)
+	}
+
+	if restarts != 1 {
+		t.Fatalf("Handler.Restart called %d times, want 1 — the restart task was swallowed "+
+			"by the duplicate-claim guard, so the forward would never come back", restarts)
+	}
+	if claims != 1 {
+		t.Fatalf("Handler.Claim called %d times, want 1 (the original claim only); "+
+			"a restart must not re-run the claim bookkeeping", claims)
+	}
+	// Re-enabling must not create a second entry.
+	if n := len(eng.state.TopologyList()); n != entryCount {
+		t.Fatalf("restart created a duplicate topology entry: %d -> %d", entryCount, n)
+	}
+	if d, _ := eng.state.TopologyDisabled("web", "frps1", 18081); d {
+		t.Fatal("entry is still disabled after the restart task was applied")
+	}
+	var sawStart bool
+	for _, e := range eng.Log.Snapshot() {
+		if e.Kind == LogForwardStart {
+			sawStart = true
+		}
+	}
+	if !sawStart {
+		t.Fatalf("log missing forward.start: %+v", eng.Log.Snapshot())
+	}
+}
+
+// TestRestartFlagSurvivesTokenSerialization: the restart task travels to the
+// owner inside the token, so the flag must round-trip through JSON. Without the
+// struct tag it would silently deserialize as false and the owner would treat
+// it as an ordinary claim — which the duplicate guard then discards.
+func TestRestartFlagSurvivesTokenSerialization(t *testing.T) {
+	eng := newTestEngine("n1", true)
+	eng.SubmitRestart(store.Local{Name: "web"}, store.Remote{Name: "frps1"},
+		store.Link{RemotePort: 18081})
+
+	blob, err := json.Marshal(&Token{Cycle: 7, State: eng.state})
+	if err != nil {
+		t.Fatal(err)
+	}
+	var back Token
+	if err := json.Unmarshal(blob, &back); err != nil {
+		t.Fatal(err)
+	}
+	var found *Task
+	for _, tk := range back.State.PendingList() {
+		if tk.Local.Name == "web" && tk.Link.RemotePort == 18081 {
+			found = tk
+		}
+	}
+	if found == nil {
+		t.Fatal("restart task did not survive token serialization")
+	}
+	if !found.Restart {
+		t.Fatal("the restart flag was lost in JSON — the owner would see a plain claim " +
+			"and the duplicate guard would discard it")
+	}
+	// Revoke and Restart must stay distinguishable.
+	if found.Revoke {
+		t.Fatal("a restart task must not also read as a revocation")
+	}
+}
+
+// TestRestartTaskReachesNonLocalOwner.
+func TestRestartOnlyAppliedByOwner(t *testing.T) {
+	restarts := 0
+	h := &fakeHandler{load: Load{MemPct: 5, NetPct: 5},
+		restart: func(ctx context.Context, tk *Task) error { restarts++; return nil }}
+
+	// Owner node claims the forward.
+	owner := NewEngine("n1", "n1:7500", "u", "p", "0.71.0", nil, h,
+		func(ctx context.Context, next string, tk *Token) error { return nil }, "n1:7500", true, "")
+	owner.state.AddPending(store.Local{Name: "web"}, store.Remote{Name: "frps1"}, store.Link{RemotePort: 18081})
+	if _, err := owner.OnToken(context.Background(), &Token{Cycle: 1, State: owner.state, SentAt: 1}); err != nil {
+		t.Fatal(err)
+	}
+	if len(owner.state.TopologyList()) != 1 {
+		t.Fatalf("precondition: owner should hold the entry, got %+v", owner.state.TopologyList())
+	}
+
+	// A peer adopts the same ring state but does not own the forward.
+	peer := NewEngine("n2", "n2:7500", "u", "p", "0.71.0", nil, h,
+		func(ctx context.Context, next string, tk *Token) error { return nil }, "n2:7500", false, "")
+	peer.AdoptState(owner.state)
+	tk := peer.SubmitRestart(store.Local{Name: "web"}, store.Remote{Name: "frps1"},
+		store.Link{RemotePort: 18081})
+	if tk == nil {
+		t.Fatal("SubmitRestart returned nil")
+	}
+
+	// The peer processes one token: it must NOT apply the restart.
+	if _, err := peer.OnToken(context.Background(), &Token{Cycle: 2, State: peer.state, SentAt: 2}); err != nil {
+		t.Fatal(err)
+	}
+	if restarts != 0 {
+		t.Fatalf("a NON-owner applied the restart %d time(s) — it would spawn an orphaned "+
+			"worker for a forward owned by somebody else", restarts)
+	}
+}
+
+// TestRestartTaskReachesNonLocalOwner: a restart published on a node that is NOT
+// the owner must survive the token round-trip and be applied by the owner.
+//
+// The naive "if not owner then continue" implementation CONSUMED the task
+// (ClaimPending had already removed it), so the owner never received it and the
+// forward stayed stopped with nothing logged as an error. Deferral must
+// re-queue it.
+//
+// Each OnToken is fed a FRESH token (as the real ring does — the successor's
+// OnToken receives the state the predecessor returned), so the deferral is
+// exercised once per hop rather than re-processing one snapshot.
+func TestRestartTaskReachesNonLocalOwner(t *testing.T) {
+	ownerApplied, submitterApplied := 0, 0
+	ownerEng := NewEngine("n2", "n2:7500", "u", "p", "0.71.0", nil,
+		&fakeHandler{load: Load{MemPct: 5, NetPct: 5},
+			restart: func(ctx context.Context, tk *Task) error { ownerApplied++; return nil }},
+		func(ctx context.Context, next string, tk *Token) error { return nil }, "n2:7500", false, "")
+
+	ownerEng.state.AddPending(store.Local{Name: "web"}, store.Remote{Name: "frps1"}, store.Link{RemotePort: 18081})
+	if _, err := ownerEng.OnToken(context.Background(), &Token{Cycle: 1, State: ownerEng.state}); err != nil {
+		t.Fatal(err)
+	}
+	if len(ownerEng.state.TopologyList()) != 1 {
+		t.Fatalf("precondition: n2 should own it, got %+v", ownerEng.state.TopologyList())
+	}
+
+	subEng := NewEngine("n1", "n1:7500", "u", "p", "0.71.0", nil,
+		&fakeHandler{load: Load{MemPct: 5, NetPct: 5},
+			restart: func(ctx context.Context, tk *Task) error { submitterApplied++; return nil }},
+		func(ctx context.Context, next string, tk *Token) error { return nil }, "n1:7500", true, "")
+	subEng.AdoptState(ownerEng.state)
+
+	subEng.SubmitRestart(store.Local{Name: "web"}, store.Remote{Name: "frps1"},
+		store.Link{RemotePort: 18081})
+
+	// One hop through the non-owner: it must defer, not apply and not drop.
+	out, err := subEng.OnToken(context.Background(), &Token{Cycle: 2, State: subEng.state, SentAt: 1})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if submitterApplied != 0 {
+		t.Fatal("the non-owner applied a restart for a forward it does not own")
+	}
+	var carried *Task
+	for _, tk := range out.State.PendingList() {
+		if tk.Restart && tk.Local.Name == "web" {
+			carried = tk
+		}
+	}
+	if carried == nil {
+		t.Fatal("the non-owner CONSUMED the restart task; the owner would never receive it")
+	}
+
+	// The owner applies the carried task.
+	if _, err := ownerEng.OnToken(context.Background(), &Token{Cycle: 3, State: out.State, SentAt: 2}); err != nil {
+		t.Fatal(err)
+	}
+	if ownerApplied != 1 {
+		t.Fatalf("the owner applied the restart %d times, want 1", ownerApplied)
+	}
 }

internal/cluster/ring_splitbrain_test.go

@@ -0,0 +1,190 @@
+package cluster
+
+import (
+	"context"
+	"encoding/json"
+	"testing"
+
+	"webui4frpc/internal/store"
+)
+
+// ringOf builds a 3-node ring where n1 is the leader.
+func ringOf(t *testing.T) (n1, n2, n3 *Engine) {
+	mk := func(id string) *Engine {
+		return NewEngine(id, id+":7500", "u", "p", "0.1.2", nil,
+			&fakeHandler{load: Load{MemPct: 5, NetPct: 5}},
+			func(ctx context.Context, next string, tk *Token) error { return nil },
+			id+":7500", true, "")
+	}
+	n1, n2, n3 = mk("n1"), mk("n2"), mk("n3")
+	seed := State{
+		LeaderID: "n1",
+		Nodes: []Node{
+			{ID: "n1", Addr: "n1:7500", Alive: true, IsLeader: true},
+			{ID: "n2", Addr: "n2:7500", Alive: true},
+			{ID: "n3", Addr: "n3:7500", Alive: true},
+		},
+		Topology:     map[string]*TopoEntry{},
+		PendingTasks: map[string]*Task{},
+	}
+	for _, e := range []*Engine{n1, n2, n3} {
+		// Round-trip through JSON so each engine holds its OWN Nodes slice and
+		// maps. AdoptState assigns the state by value, but Nodes is a slice, so
+		// handing all three the same seed would alias one backing array and make
+		// a mutation on one node visible on the others — an artefact of the test
+		// harness, not of the real transport (each process deserializes its own
+		// copy from the token).
+		blob, err := json.Marshal(seed)
+		if err != nil {
+			t.Fatal(err)
+		}
+		var own State
+		if err := json.Unmarshal(blob, &own); err != nil {
+			t.Fatal(err)
+		}
+		e.AdoptState(own)
+	}
+	return
+}
+
+// TestBecomeLeaderDeclinesToLiveLeader is the regression test for the live split
+// brain.
+//
+// Leadership used to be self-declared: any node that failed to reach the leader
+// it believed in promoted itself immediately. Both sides of a partition take that
+// path, so two rings ran concurrently and each emptied the other's topology into
+// pending — observed live as two nodes naming the same leader while marking each
+// other dead, with pending=3 and topology=0 on both.
+//
+// A node must therefore consult the ring before promoting itself.
+func TestBecomeLeaderDeclinesToLiveLeader(t *testing.T) {
+	_, n2, _ := ringOf(t)
+	// n3 has already promoted itself and n2 has learned that view: n1 is dead,
+	// n3 is alive and leading. n2 cannot reach n3 (that is why it is asking).
+	for i := range n2.state.Nodes {
+		switch n2.state.Nodes[i].ID {
+		case "n1":
+			n2.state.Nodes[i].Alive = false
+			n2.state.Nodes[i].IsLeader = false
+		case "n3":
+			n2.state.Nodes[i].Alive = true
+			n2.state.Nodes[i].IsLeader = true
+		}
+	}
+	n2.state.LeaderID = "n3"
+
+	if n2.becomeLeader() {
+		t.Fatal("n2 promoted itself while n3 was alive and leading in its view — " +
+			"two leaders is exactly the split brain this guards against")
+	}
+	if n2.state.LeaderID == "n2" {
+		t.Fatal("LeaderID was overwritten despite declining")
+	}
+	if n2.state.Nodes[0].IsLeader && n2.state.Nodes[0].ID == "n2" {
+		t.Fatal("n2 marked itself leader despite declining")
+	}
+}
+
+// TestBecomeLeaderPromotesWhenSoleSurvivor: the guard must not break genuine
+// failover. With every other node dead, promotion has to happen.
+func TestBecomeLeaderPromotesWhenSoleSurvivor(t *testing.T) {
+	_, n2, _ := ringOf(t)
+	for _, id := range []string{"n1", "n3"} {
+		n2.state.MarkOffline(id)
+	}
+	if !n2.becomeLeader() {
+		t.Fatal("n2 declined promotion with every other node dead — that breaks failover")
+	}
+	if n2.state.LeaderID != "n2" {
+		t.Fatalf("leader=%q, want n2", n2.state.LeaderID)
+	}
+}
+
+// TestTakeoverLeavesLiveLeadersForwardsAlone pins the second half of the fix.
+//
+// When a heartbeat fails, the node used to mark the leader dead and re-queue ALL
+// of that leader's forwards unconditionally — even when the ring still showed a
+// live leader, i.e. this node was merely the partitioned side. That is what
+// emptied a working ring: both sides ended up with topology=0 and pending=3.
+//
+// Now the re-hang happens only if we actually take leadership.
+func TestTakeoverLeavesLiveLeadersForwardsAlone(t *testing.T) {
+	_, n2, _ := ringOf(t)
+	// n1 (the leader n2 believes in) owns a forward.
+	n2.state.AddTopology(&Task{
+		ID:    "t1",
+		Local: store.Local{Name: "web"}, Remote: store.Remote{Name: "frps1"},
+		Link: store.Link{RemotePort: 18081},
+	}, "n1")
+	before := len(n2.state.TopologyList())
+	pendingBefore := len(n2.state.PendingList())
+
+	// n2 still believes n1 is the leader and cannot reach it, so it tries to take
+	// over. But its state ALSO lists n3 as alive-and-leader — the view it learned
+	// when n3 promoted itself. Two claims, one ring: n2 is the partitioned side
+	// and must stand down.
+	for i := range n2.state.Nodes {
+		switch n2.state.Nodes[i].ID {
+		case "n1":
+			n2.state.Nodes[i].Alive = true // still marked alive; MarkOffline happens inside
+			n2.state.Nodes[i].IsLeader = true
+		case "n3":
+			n2.state.Nodes[i].Alive = true
+			n2.state.Nodes[i].IsLeader = true
+		}
+	}
+	n2.state.LeaderID = "n1"
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	n2.takeOverFromDeadLeader(ctx)
+
+	if n2.state.LeaderID == "n2" {
+		t.Error("n2 took over while a live leader existed")
+	}
+	if after := len(n2.state.TopologyList()); after != before {
+		t.Errorf("a declined takeover still re-queued the live leader's forwards: topology %d -> %d",
+			before, after)
+	}
+	if p := len(n2.state.PendingList()); p != pendingBefore {
+		t.Errorf("a declined takeover still pushed work into pending: %d -> %d", pendingBefore, p)
+	}
+}
+
+// TestTakeoverReassignsWhenSoleSurvivor is the other half: genuine failover must
+// still work, otherwise the guard would have broken the design's promise that a
+// dead leader is replaced and its forwards re-hung.
+func TestTakeoverReassignsWhenSoleSurvivor(t *testing.T) {
+	_, n2, _ := ringOf(t)
+	n2.state.AddTopology(&Task{
+		ID:    "t1",
+		Local: store.Local{Name: "web"}, Remote: store.Remote{Name: "frps1"},
+		Link: store.Link{RemotePort: 18081},
+	}, "n1") // owned by the leader n1
+
+	// n1 is the leader and is dead; n3 is dead too. n2 is the sole survivor.
+	for _, id := range []string{"n1", "n3"} {
+		n2.state.MarkOffline(id)
+	}
+	n2.state.Nodes[0].IsLeader = false // n1 no longer claims leadership
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	n2.takeOverFromDeadLeader(ctx)
+
+	if n2.state.LeaderID != "n2" {
+		t.Fatalf("n2 did not take over as the sole survivor, leader=%q", n2.state.LeaderID)
+	}
+	if len(n2.state.TopologyList()) != 0 {
+		t.Error("the dead leader's forwards should have been re-hung")
+	}
+	found := false
+	for _, tk := range n2.state.PendingList() {
+		if tk.Local.Name == "web" {
+			found = true
+		}
+	}
+	if !found {
+		t.Error("the dead leader's forward was not re-queued for another node")
+	}
+}

internal/cluster/token_transport.go

@@ -8,7 +8,6 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
-	"log"
 	"net/http"
 	"time"
 )
@@ -52,11 +51,16 @@ func (t *TokenTransport) SendTo(getAddr func(nodeID string) string) func(ctx con
 		start := time.Now()
 		resp, err := cli.Do(req)
 		if err != nil {
-			log.Printf("token-send %s: size=%dB elapsed=%v err=%v", addr, len(body), time.Since(start).Round(time.Millisecond), err)
+			// NOT demoted: a failed send is the "neighbor offline" signal the
+			// ring's fault paths are diagnosed from.
+			logf("token-send %s: size=%dB elapsed=%v err=%v", addr, len(body), time.Since(start).Round(time.Millisecond), err)
 			return err
 		}
 		defer resp.Body.Close()
-		log.Printf("token-send %s: size=%dB elapsed=%v -> %d", addr, len(body), time.Since(start).Round(time.Millisecond), resp.StatusCode)
+		// Success path is a per-round heartbeat (size/elapsed/200 repeat
+		// verbatim every cycle) — debug only. The status-code check below stays
+		// loud on purpose, so a non-2xx still shows up without the debug flag.
+		debugToken("token-send %s: size=%dB elapsed=%v -> %d", addr, len(body), time.Since(start).Round(time.Millisecond), resp.StatusCode)
 		if resp.StatusCode >= 400 {
 			return fmt.Errorf("token POST %s -> HTTP %d", url, resp.StatusCode)
 		}

internal/httpapi/forward_revoke_test.go

@@ -0,0 +1,204 @@
+package httpapi
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"path/filepath"
+	"testing"
+
+	"webui4frpc/internal/cluster"
+	"webui4frpc/internal/process"
+	"webui4frpc/internal/store"
+)
+
+// newRingTestHandler builds a Handler WITH a ring engine attached, so the
+// paths that publish tasks into the token actually execute. newTestHandler
+// leaves Ring nil, which silently skips them — a test built on it can pass
+// while the publish side is completely broken.
+func newRingTestHandler(t *testing.T) (*Handler, *cluster.Engine, *httptest.Server) {
+	t.Helper()
+	dir := t.TempDir()
+	st, err := store.New(filepath.Join(dir, "test.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { _ = st.Close() })
+
+	pm := process.NewManager(process.Options{
+		ConfigsDir:      filepath.Join(dir, "configs"),
+		LogsDir:         filepath.Join(dir, "logs"),
+		BinaryPath:      func() string { return "" },
+		Render:          func(string) ([]byte, error) { return []byte(`{}`), nil },
+		AutoRestart:     func(string) bool { return false },
+		RestartInterval: func() int { return 5 },
+	})
+	ring := cluster.NewEngine("n1", "n1:7500", "u", "p", "0.1.0", nil,
+		&cluster.AppHandler{},
+		func(ctx context.Context, next string, tk *cluster.Token) error { return nil },
+		"n1:7500", true, "")
+	h := &Handler{Store: st, Process: pm, WorkDir: dir, User: "admin", Password: "pw", Ring: ring}
+	mux, err := NewServeMux(h)
+	if err != nil {
+		t.Fatal(err)
+	}
+	ts := httptest.NewServer(mux)
+	t.Cleanup(ts.Close)
+	return h, ring, ts
+}
+
+func saveCanvas(t *testing.T, srv *httptest.Server, body string) {
+	t.Helper()
+	req, _ := http.NewRequest(http.MethodPut, srv.URL+"/api/manager/canvas", bytes.NewBufferString(body))
+	req.SetBasicAuth("admin", "pw")
+	resp, err := srv.Client().Do(req)
+	if err != nil {
+		t.Fatal(err)
+	}
+	resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("canvas save status = %d", resp.StatusCode)
+	}
+}
+
+// TestStopForwardPublishesDisabledFlagInRevokeTask is the regression test for
+// the actual defect.
+//
+// stopForward() read the link, called SetLinkDisabled(true), and then handed
+// the STALE copy (disabled=false) to RevokeTask. The revoke travels to the
+// node that OWNS the forward, and that node's Claim/Revoke path keys off the
+// flag — so a stale false meant:
+//   - the owner could not tell the stop was deliberate, and
+//   - nothing retired the topology entry,
+//
+// so the next restart's reconcile re-claimed the forward and spawned a worker
+// for something the user had stopped (seen live: endless connection-refused
+// against an intentionally-down service).
+//
+// This asserts the flag ON THE PUBLISHED TASK, which is the value that was
+// actually wrong. It cannot be satisfied by the store write alone.
+func TestStopForwardPublishesDisabledFlagInRevokeTask(t *testing.T) {
+	_, ring, ts := newRingTestHandler(t)
+
+	saveCanvas(t, ts, `{
+		"locals": [{"name":"svc","ip":"127.0.0.1","port":59999,"protocol":"tcp"}],
+		"remotes": [{"name":"srv-a","ip":"1.2.3.4","port":7000,"enabled":true}],
+		"links": [{"local":"svc","remote":"srv-a","remotePort":45999}]
+	}`)
+
+	// Stop the forward over the API.
+	b, _ := json.Marshal(stopForwardReq{"svc", "srv-a", 45999})
+	req, _ := http.NewRequest(http.MethodPost, ts.URL+"/api/manager/forwards/stop", bytes.NewReader(b))
+	req.SetBasicAuth("admin", "pw")
+	resp, err := ts.Client().Do(req)
+	if err != nil {
+		t.Fatal(err)
+	}
+	resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("stop status = %d", resp.StatusCode)
+	}
+
+	// Find the revoke task that was published into the token.
+	var revoke *cluster.Task
+	for _, tk := range ring.State().PendingList() {
+		if tk.Revoke && tk.Local.Name == "svc" && tk.Link.RemotePort == 45999 {
+			revoke = tk
+			break
+		}
+	}
+	if revoke == nil {
+		t.Fatal("stop did not publish a revoke task for the forward")
+	}
+	if !revoke.Link.Disabled {
+		t.Fatal("the published revoke task carries disabled=false — the owner node cannot tell " +
+			"this stop was deliberate, which is the bug that let stopped forwards resurrect")
+	}
+}
+
+// TestStopThenStartPublishesRestartTask guards the failure mode that
+// mark-don't-remove introduces, and that only shows up on the WIRE.
+//
+// A stop keeps the topology entry (it is the carrier for the flag), so on
+// re-enable:
+//   - SubmitTask dedupes, because the entry exists;
+//   - the claim path discards any task for a forward that already has an owner.
+//
+// Both channels therefore refuse, no task is published, the owner never learns
+// about the re-enable, and the forward stays stopped forever — a forward the
+// user can stop but never restart.
+//
+// Asserting the task is PUBLISHED is the point: asserting only that the flag
+// cleared would pass while nothing reached the owner. (The earlier version of
+// this test did exactly that and stayed green against the broken code.)
+func TestStopThenStartPublishesRestartTask(t *testing.T) {
+	_, ring, ts := newRingTestHandler(t)
+
+	saveCanvas(t, ts, `{
+		"locals": [{"name":"svc","ip":"127.0.0.1","port":59999,"protocol":"tcp"}],
+		"remotes": [{"name":"srv-a","ip":"1.2.3.4","port":7000,"enabled":true}],
+		"links": [{"local":"svc","remote":"srv-a","remotePort":45999}]
+	}`)
+	// Claim it so a topology entry exists (that is what makes the two normal
+	// channels refuse later).
+	if _, err := ring.OnToken(context.Background(), &cluster.Token{Cycle: 1, State: *ring.State()}); err != nil {
+		t.Fatal(err)
+	}
+	if !ring.HasActiveTask("svc", "srv-a", 45999) {
+		t.Fatalf("precondition: forward should be active, topo=%+v", ring.State().TopologyList())
+	}
+
+	forwards := func(action string) {
+		t.Helper()
+		b, _ := json.Marshal(stopForwardReq{"svc", "srv-a", 45999})
+		req, _ := http.NewRequest(http.MethodPost, ts.URL+"/api/manager/forwards/"+action, bytes.NewReader(b))
+		req.SetBasicAuth("admin", "pw")
+		resp, err := ts.Client().Do(req)
+		if err != nil {
+			t.Fatal(err)
+		}
+		resp.Body.Close()
+		if resp.StatusCode != http.StatusOK {
+			t.Fatalf("%s status = %d", action, resp.StatusCode)
+		}
+	}
+
+	// --- stop: entry kept, marked disabled -------------------------------
+	forwards("stop")
+	if d, known := ring.TopologyDisabled("svc", "srv-a", 45999); !known || !d {
+		t.Fatalf("stop did not mark the topology entry disabled (known=%v disabled=%v)", known, d)
+	}
+	if ring.HasActiveTask("svc", "srv-a", 45999) {
+		t.Fatal("a stopped forward must not count as active")
+	}
+
+	// --- start: a task MUST reach the owner ------------------------------
+	forwards("start")
+	if d, _ := ring.TopologyDisabled("svc", "srv-a", 45999); d {
+		t.Fatal("start did not clear the disabled flag")
+	}
+	if !ring.HasActiveTask("svc", "srv-a", 45999) {
+		t.Fatal("after start the forward must be active again")
+	}
+
+	// The actual regression: something has to be queued for the owner. The
+	// re-enable must be published as a restart task, since a plain creation task
+	// would be deduped or discarded.
+	var restart *cluster.Task
+	for _, tk := range ring.State().PendingList() {
+		if tk.Restart && tk.Local.Name == "svc" && tk.Link.RemotePort == 45999 {
+			restart = tk
+			break
+		}
+	}
+	if restart == nil {
+		t.Fatalf("start published no restart task — the owner would never bring the "+
+			"worker back, leaving the forward permanently stopped (pending=%+v)",
+			ring.State().PendingList())
+	}
+	if restart.Link.Disabled {
+		t.Fatal("the restart task carries disabled=true, so the owner would re-apply the stop")
+	}
+}

internal/httpapi/forward_stop_test.go

@@ -0,0 +1,141 @@
+package httpapi
+
+import (
+	"bytes"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+)
+
+// stopForwardReq mirrors the forwards start/stop request body.
+type stopForwardReq struct {
+	Local      string `json:"local"`
+	Remote     string `json:"remote"`
+	RemotePort int    `json:"remotePort"`
+}
+
+func postForwards(t *testing.T, srv *httptest.Server, action string, body stopForwardReq) int {
+	t.Helper()
+	b, _ := json.Marshal(body)
+	req, _ := http.NewRequest(http.MethodPost, srv.URL+"/api/manager/forwards/"+action, bytes.NewReader(b))
+	req.SetBasicAuth("admin", "pw")
+	req.Header.Set("Content-Type", "application/json")
+	resp, err := srv.Client().Do(req)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	return resp.StatusCode
+}
+
+// TestStopForwardPersistsDisabledFlag is the regression test for the
+// "stopped forwards resurrect on restart" bug.
+//
+// The original defect: stopForward() read the link, called
+// SetLinkDisabled(true), but then handed the STALE (disabled=false) copy to
+// RevokeTask — so the disabled flag never reached the node owning the forward,
+// and nothing removed the topology entry. On the next restart the startup
+// reconcile saw an owned forward with no worker and re-claimed it, spawning a
+// worker for a forward the user had deliberately stopped (observed live:
+// ~14k "proxy already exists" retries and endless connection-refused against a
+// service that was intentionally down).
+//
+// The contract this pins: after a successful stop, the persisted link MUST be
+// disabled — that flag is the single source of truth the claim path consults.
+func TestStopForwardPersistsDisabledFlag(t *testing.T) {
+	h, ts := newTestHandler(t)
+
+	body := `{
+		"locals": [{"name":"svc","ip":"127.0.0.1","port":59999,"protocol":"tcp"}],
+		"remotes": [{"name":"srv-a","ip":"1.2.3.4","port":7000,"enabled":true}],
+		"links": [{"local":"svc","remote":"srv-a","remotePort":45999}]
+	}`
+	req, _ := http.NewRequest(http.MethodPut, ts.URL+"/api/manager/canvas", bytes.NewBufferString(body))
+	req.SetBasicAuth("admin", "pw")
+	resp, err := ts.Client().Do(req)
+	if err != nil {
+		t.Fatal(err)
+	}
+	resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("canvas save status = %d", resp.StatusCode)
+	}
+
+	// The forward starts enabled.
+	ln, found, err := h.Store.LinkByTriple("svc", "srv-a", 45999)
+	if err != nil || !found {
+		t.Fatalf("link not persisted: found=%v err=%v", found, err)
+	}
+	if ln.Disabled {
+		t.Fatal("a freshly saved forward must start enabled")
+	}
+
+	// Stop it.
+	if code := postForwards(t, ts, "stop", stopForwardReq{"svc", "srv-a", 45999}); code != http.StatusOK {
+		t.Fatalf("stop status = %d, want 200", code)
+	}
+
+	// Persisted flag must now be set — this is what the claim path reads.
+	ln, found, err = h.Store.LinkByTriple("svc", "srv-a", 45999)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !found {
+		t.Fatal("link vanished after stop; stop must be non-destructive")
+	}
+	if !ln.Disabled {
+		t.Fatal("stop did not persist disabled=true — the startup reconcile would resurrect this forward")
+	}
+
+	// Start must clear it again (the user-facing re-enable path).
+	if code := postForwards(t, ts, "start", stopForwardReq{"svc", "srv-a", 45999}); code != http.StatusOK {
+		t.Fatalf("start status = %d, want 200", code)
+	}
+	ln, _, err = h.Store.LinkByTriple("svc", "srv-a", 45999)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if ln.Disabled {
+		t.Fatal("start did not clear disabled; a re-enabled forward would stay stopped")
+	}
+}
+
+// TestStopForwardIsNonDestructive pins the per-forward stop semantics the
+// revoke path was specifically rewritten for: stopping one forward must not
+// touch a sibling forward that shares the same local or remote.
+func TestStopForwardIsNonDestructive(t *testing.T) {
+	h, ts := newTestHandler(t)
+
+	body := `{
+		"locals": [{"name":"svc","ip":"127.0.0.1","port":59999,"protocol":"tcp"}],
+		"remotes": [{"name":"srv-a","ip":"1.2.3.4","port":7000,"enabled":true}],
+		"links": [
+			{"local":"svc","remote":"srv-a","remotePort":45999},
+			{"local":"svc","remote":"srv-a","remotePort":46000}
+		]
+	}`
+	req, _ := http.NewRequest(http.MethodPut, ts.URL+"/api/manager/canvas", bytes.NewBufferString(body))
+	req.SetBasicAuth("admin", "pw")
+	resp, err := ts.Client().Do(req)
+	if err != nil {
+		t.Fatal(err)
+	}
+	resp.Body.Close()
+
+	if code := postForwards(t, ts, "stop", stopForwardReq{"svc", "srv-a", 45999}); code != http.StatusOK {
+		t.Fatalf("stop status = %d", code)
+	}
+
+	stopped, _, _ := h.Store.LinkByTriple("svc", "srv-a", 45999)
+	sibling, found, _ := h.Store.LinkByTriple("svc", "srv-a", 46000)
+	if !found {
+		t.Fatal("sibling forward was destroyed by stopping its neighbour")
+	}
+	if !stopped.Disabled {
+		t.Error("the stopped forward should be disabled")
+	}
+	if sibling.Disabled {
+		t.Error("the sibling forward must stay enabled — per-forward stop, not per-local/remote")
+	}
+}

internal/httpapi/handlers.go

@@ -245,7 +245,12 @@ func (h *Handler) applyCanvas(w http.ResponseWriter, r *http.Request, canvas *ca
 			}
 			if ln.Disabled {
 				// Stopped on the forwards page: make sure it leaves the topology.
-				if h.Ring.HasTask(ln.Local, ln.Remote, ln.RemotePort) {
+				// Guard on an ACTIVE task — an entry that is already marked
+				// disabled has nothing left to revoke, and re-issuing a revocation
+				// for it on every canvas save would be pure noise. (HasTask, which
+				// also matches disabled entries, is the right predicate for the
+				// "already handled" question this branch is not asking.)
+				if h.Ring.HasActiveTask(ln.Local, ln.Remote, ln.RemotePort) {
 					h.Ring.RevokeTask(loc, rem, ln)
 				}
 				continue

internal/httpapi/handlers_forwards.go

@@ -84,6 +84,25 @@ func (h *Handler) startForward(local, remote string, port int) error {
 		}
 		return nil
 	}
+	if h.Ring != nil {
+		// Publish the enable into the topology so it rides the ring (a peer that
+		// still holds the stale "stopped" copy learns about it on adoption).
+		h.Ring.UpdateTopologyDisabled(local, remote, port, false)
+	}
+	// A cluster forward whose topology entry still exists needs its OWNER to
+	// bring the worker back, and neither of the normal channels can do it:
+	//   - SubmitTask is idempotency-guarded, and the entry still exists (a stop
+	//     keeps it as the flag's carrier), so the submission is deduped away;
+	//   - the claim path drops any task for a forward that already has an
+	//     owner, as a defence against duplicate-claim collisions.
+	// So a re-enable of an existing entry is published as a dedicated RESTART
+	// task, which the owner applies unconditionally. Without it a forward the
+	// user can stop but not restart — which is what marking-instead-of-removing
+	// would otherwise have produced.
+	if h.Ring != nil && h.Ring.HasTopologyEntry(local, remote, port) {
+		h.Ring.SubmitRestart(loc, rem, ln)
+		return nil
+	}
 	if h.Ring != nil {
 		h.Ring.SubmitTask(loc, rem, ln)
 	}
@@ -109,6 +128,24 @@ func (h *Handler) stopForward(local, remote string, port int) error {
 	} else {
 		_ = h.Store.SetLinkDisabled(local, remote, port, true)
 	}
+	// The revoke task travels to whichever node OWNS the forward, and that node
+	// re-reads the disabled flag from its own store before starting a worker —
+	// so the flag has to be set on every node that has a copy of this link, not
+	// just the one handling this request. Propagating Disabled on the task lets
+	// the owner's RevokeFn stop the worker even if its own store row is stale.
+	//
+	// This also fixes a latent inconsistency: `ln` was read BEFORE the
+	// SetLinkDisabled(true) above, so the link published into the token still
+	// carried disabled=false and got copied into the topology entry verbatim.
+	ln.Disabled = true
+	// Publish the stop into the topology BEFORE revoking: the revoke retires the
+	// own-side worker/entry, so the flag must already exist somewhere that
+	// survives it and travels the ring. This is the send half of cluster-wide
+	// disabled propagation (see store.ReconcileLinkDisabled for the receive
+	// half).
+	if h.Ring != nil {
+		h.Ring.UpdateTopologyDisabled(local, remote, port, true)
+	}
 	if loc.LocalOnly {
 		if h.Process != nil {
 			key := process.WorkerKey(local, remote, port)

internal/httpapi/server.go

@@ -48,6 +48,8 @@ type Handler struct {
 	Cluster *cluster.Registry
 	// Ring is the token-ring engine (M6). Nil disables ring routes.
 	Ring *cluster.Engine
+	// Balancer is optional: when nil the balance endpoint reports 404.
+	Balancer *cluster.Balancer
 	// SelfAddr is this node's reachable listen address (from -addr), used for
 	// cluster discovery so peers can reach back for binary exchange.
 	SelfAddr string
@@ -114,6 +116,8 @@ func NewServeMux(h *Handler) (http.Handler, error) {
 	mux.HandleFunc(apiPrefix+"/cluster/nodes", h.auth("read")(h.handleClusterNodes))
 	mux.HandleFunc(apiPrefix+"/cluster/cache", h.auth("read")(h.handleClusterCache)) // GET read; POST write-checked inside
 	mux.HandleFunc(apiPrefix+"/cluster/ring", h.auth("read")(h.handleClusterRing))
+	mux.HandleFunc(apiPrefix+"/cluster/balance", h.auth("read")(h.handleClusterBalance))
+	mux.HandleFunc(apiPrefix+"/cluster/balance/mode", h.auth("write")(h.handleClusterBalanceMode))
 	mux.HandleFunc(apiPrefix+"/node/logs", h.auth("read")(h.handleNodeLogs))
 	mux.HandleFunc(apiPrefix+"/cluster/logs/export", h.auth("read")(h.handleClusterLogsExport))
 
@@ -634,6 +638,64 @@ func (h *Handler) handleClusterRing(w http.ResponseWriter, r *http.Request) {
 	writeJSON(w, http.StatusOK, h.Ring.Snapshot())
 }
 
+// handleClusterBalance reports the load balancer's read-only view: per-node
+// scores, whether anything is over the threshold, and the migration it would
+// perform. Publishing nothing is the point — an operator asking "is this cluster
+// balanced?" currently gets silence from the log in the healthy case, which is
+// indistinguishable from balancing being switched off.
+func (h *Handler) handleClusterBalance(w http.ResponseWriter, r *http.Request) {
+	if r.Method != http.MethodGet {
+		methodNotAllowed(w)
+		return
+	}
+	if h.Ring == nil {
+		http.Error(w, "ring engine not enabled", http.StatusNotFound)
+		return
+	}
+	if h.Balancer == nil {
+		http.Error(w, "load balancer not enabled", http.StatusNotFound)
+		return
+	}
+	rep := h.Balancer.Report()
+	// CanControl is decided by the caller's own level rather than by the
+	// presence of a stored value, so a read-only principal is never shown a
+	// switch that would 403.
+	rep.CanControl = hasLevel(r, "write")
+	writeJSON(w, http.StatusOK, rep)
+}
+
+// handleClusterBalanceMode switches load balancing on/off at runtime.
+//
+// It is per-node on purpose. Every member decides about its OWN forwards, so the
+// switch belongs to each node: an operator may well want to turn releasing on
+// everywhere, but the server should not pretend those are one atomic action.
+// (A cluster-wide control can be layered on top later; the primitive here is the
+// one that actually has an owner.)
+//
+// Write-level auth: releasing stops a worker and drops a live connection.
+func (h *Handler) handleClusterBalanceMode(w http.ResponseWriter, r *http.Request) {
+	if r.Method != http.MethodPost {
+		methodNotAllowed(w)
+		return
+	}
+	if h.Balancer == nil {
+		http.Error(w, "load balancer not enabled", http.StatusNotFound)
+		return
+	}
+	var req struct {
+		Mode string `json:"mode"`
+	}
+	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+		http.Error(w, "parse json: "+err.Error(), http.StatusBadRequest)
+		return
+	}
+	if err := h.Balancer.SetModeFromUI(req.Mode); err != nil {
+		http.Error(w, err.Error(), http.StatusBadRequest)
+		return
+	}
+	writeJSON(w, http.StatusOK, h.Balancer.Report())
+}
+
 // handleClusterJoin accepts a newcomer join request: the target node inserts
 // the newcomer after itself in the ring and returns the updated ring state
 // for the newcomer to adopt.

internal/store/link_test.go

@@ -0,0 +1,288 @@
+package store
+
+import (
+	"path/filepath"
+	"testing"
+)
+
+// seed inserts the local + remote rows a link's foreign keys require.
+// (links.local / links.remote reference their own tables, so a link cannot
+// exist on its own — the same reason ClaimFn upserts them before ReplaceLinks.)
+func seed(t *testing.T, st *Store, locals []string, remote string) {
+	t.Helper()
+	for _, n := range locals {
+		if err := st.UpsertLocal(Local{Name: n, IP: "127.0.0.1", Port: 8080, Protocol: "tcp"}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if err := st.UpsertRemote(Remote{Name: remote, IP: "1.2.3.4", Port: 7000, Token: "tok", Enabled: true}); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// TestLinkByTripleSurvivesReplaceLinks pins the reason LinkByTriple exists.
+//
+// ReplaceLinks() rewrites the whole table with DELETE + re-INSERT, so sqlite
+// hands every row a FRESH autoincrement id. A Link captured before such a
+// write (e.g. one riding inside a ring token) therefore carries an id that
+// either matches a different forward or matches nothing. The natural key
+// (local, remote, remotePort) is what every caller actually identifies a
+// forward by, and it must survive those rewrites.
+func TestLinkByTripleSurvivesReplaceLinks(t *testing.T) {
+	st, err := New(filepath.Join(t.TempDir(), "test.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer st.Close()
+
+	seed(t, st, []string{"alpha", "beta", "gamma"}, "srv")
+	links := []Link{
+		{Local: "alpha", Remote: "srv", RemotePort: 100},
+		{Local: "beta", Remote: "srv", RemotePort: 200},
+		{Local: "gamma", Remote: "srv", RemotePort: 300},
+	}
+	if err := st.ReplaceLinks(links); err != nil {
+		t.Fatal(err)
+	}
+
+	// Capture the ids as the ring would have them.
+	before := map[string]int64{}
+	all, err := st.ListLinks()
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, l := range all {
+		before[l.Local] = l.ID
+	}
+	if len(before) != 3 {
+		t.Fatalf("expected 3 links, got %d", len(before))
+	}
+
+	// Rewrite the table (this is what saveCanvas and ClaimFn both do).
+	if err := st.ReplaceLinks(links); err != nil {
+		t.Fatal(err)
+	}
+	after, err := st.ListLinks()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(after) != 3 {
+		t.Fatalf("expected 3 links after rewrite, got %d", len(after))
+	}
+
+	// The natural key must still resolve to the right forward, with its
+	// disabled flag and group intact.
+	for _, l := range after {
+		if l.Disabled {
+			t.Errorf("link %s unexpectedly disabled after a plain rewrite", l.Local)
+		}
+	}
+	got, found, err := st.LinkByTriple("beta", "srv", 200)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !found {
+		t.Fatal("LinkByTriple failed to find beta after ReplaceLinks")
+	}
+	if got.Local != "beta" || got.RemotePort != 200 {
+		t.Fatalf("LinkByTriple returned the wrong row: %+v", got)
+	}
+}
+
+// TestLinkByTripleNotFoundIsNotError documents the contract callers rely on:
+// "no persisted opinion yet" is (Link{}, false, nil), not an error. A fresh
+// claim of a link with no row must be allowed to start.
+func TestLinkByTripleNotFoundIsNotError(t *testing.T) {
+	st, err := New(filepath.Join(t.TempDir(), "test.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer st.Close()
+
+	got, found, err := st.LinkByTriple("nope", "srv", 1234)
+	if err != nil {
+		t.Fatalf("missing link must not be an error, got %v", err)
+	}
+	if found {
+		t.Fatalf("missing link reported as found: %+v", got)
+	}
+	if got.Local != "" || got.RemotePort != 0 {
+		t.Fatalf("expected zero Link on miss, got %+v", got)
+	}
+}
+
+// TestLinkByTripleReadsDisabledFlag is the store-level half of the
+// "stopped forwards resurrect on restart" bug: the claim path asks the store
+// whether the user disabled this forward, so this lookup must return the flag
+// as persisted.
+func TestLinkByTripleReadsDisabledFlag(t *testing.T) {
+	st, err := New(filepath.Join(t.TempDir(), "test.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer st.Close()
+
+	seed(t, st, []string{"mc"}, "srv")
+	if err := st.ReplaceLinks([]Link{{Local: "mc", Remote: "srv", RemotePort: 25565, Group: "game"}}); err != nil {
+		t.Fatal(err)
+	}
+	if err := st.SetLinkDisabled("mc", "srv", 25565, true); err != nil {
+		t.Fatal(err)
+	}
+
+	ln, found, err := st.LinkByTriple("mc", "srv", 25565)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !found {
+		t.Fatal("expected to find the link")
+	}
+	if !ln.Disabled {
+		t.Fatal("expected Disabled=true to be visible through LinkByTriple")
+	}
+	if ln.Group != "game" {
+		t.Fatalf("group should survive, got %q", ln.Group)
+	}
+
+	// ...and the user-facing start path must be able to clear it again.
+	if err := st.SetLinkDisabled("mc", "srv", 25565, false); err != nil {
+		t.Fatal(err)
+	}
+	if ln, _, _ := st.LinkByTriple("mc", "srv", 25565); ln.Disabled {
+		t.Fatal("expected Disabled=false after clearing")
+	}
+}
+
+// TestGetLinkByIDIsStaleAfterReplaceLinks documents WHY callers must not use
+// GetLink(id) with a previously captured id. It is not a fix — it is the trap
+// being pinned shut, so the hazard stays visible if someone reintroduces it.
+func TestGetLinkByIDIsStaleAfterReplaceLinks(t *testing.T) {
+	st, err := New(filepath.Join(t.TempDir(), "test.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer st.Close()
+
+	seed(t, st, []string{"alpha", "beta", "gamma"}, "srv")
+	if err := st.ReplaceLinks([]Link{
+		{Local: "alpha", Remote: "srv", RemotePort: 100},
+		{Local: "beta", Remote: "srv", RemotePort: 200},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	all, _ := st.ListLinks()
+	var staleID int64
+	for _, l := range all {
+		if l.Local == "alpha" {
+			staleID = l.ID
+		}
+	}
+
+	if err := st.ReplaceLinks([]Link{
+		{Local: "alpha", Remote: "srv", RemotePort: 100},
+		{Local: "beta", Remote: "srv", RemotePort: 200},
+		{Local: "gamma", Remote: "srv", RemotePort: 300},
+	}); err != nil {
+		t.Fatal(err)
+	}
+
+	// The old id may still resolve, but to whatever row now occupies that
+	// id — which is exactly the silent-mis-target hazard. Assert that the
+	// natural key remains the only safe handle.
+	if ln, ok := st.GetLink(staleID); ok && ln.Local != "alpha" {
+		t.Logf("stale id %d now points at %q (hazard confirmed; use LinkByTriple)", staleID, ln.Local)
+	}
+	if got, found, _ := st.LinkByTriple("alpha", "srv", 100); !found || got.Local != "alpha" {
+		t.Fatalf("natural key must stay reliable, got %+v found=%v", got, found)
+	}
+}
+
+// TestReconcileLinkDisabledLearnsPeerDecision is the store half of
+// cluster-wide stop propagation. A node that did NOT serve the stop request has
+// no reason to know about it, and its links table is node-local — so the flag
+// arrives via the ring and lands here. The case that matters is the OWNER of a
+// forward on a different machine: before this existed, that node's copy still
+// read "enabled", so it kept (or re-spawned) the worker for a forward the user
+// had explicitly stopped.
+func TestReconcileLinkDisabledLearnsPeerDecision(t *testing.T) {
+	st, err := New(filepath.Join(t.TempDir(), "test.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer st.Close()
+	seed(t, st, []string{"mc"}, "srv")
+	if err := st.ReplaceLinks([]Link{{Local: "mc", Remote: "srv", RemotePort: 25565, Group: "game"}}); err != nil {
+		t.Fatal(err)
+	}
+
+	// A peer's stop arrives.
+	if err := st.ReconcileLinkDisabled("mc", "srv", 25565, true); err != nil {
+		t.Fatal(err)
+	}
+	ln, found, err := st.LinkByTriple("mc", "srv", 25565)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !found {
+		t.Fatal("link disappeared during reconcile")
+	}
+	if !ln.Disabled {
+		t.Fatal("the peer's stop did not land in the local store")
+	}
+	if ln.Group != "game" {
+		t.Fatalf("reconcile must not clobber other fields, group=%q", ln.Group)
+	}
+
+	// A peer's re-enable arrives.
+	if err := st.ReconcileLinkDisabled("mc", "srv", 25565, false); err != nil {
+		t.Fatal(err)
+	}
+	if ln, _, _ := st.LinkByTriple("mc", "srv", 25565); ln.Disabled {
+		t.Fatal("the peer's re-enable did not land")
+	}
+}
+
+// TestReconcileLinkDisabledCreatesPlaceholderForUnknownForward: a node that has
+// never seen the forward still must remember that it is stopped, otherwise a
+// later claim on that node would resurrect it.
+func TestReconcileLinkDisabledCreatesPlaceholderForUnknownForward(t *testing.T) {
+	st, err := New(filepath.Join(t.TempDir(), "test.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer st.Close()
+	seed(t, st, []string{"ghost"}, "srv")
+
+	if err := st.ReconcileLinkDisabled("ghost", "srv", 9999, true); err != nil {
+		t.Fatal(err)
+	}
+	ln, found, err := st.LinkByTriple("ghost", "srv", 9999)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !found {
+		t.Fatal("a stopped-but-unknown forward must be remembered, or a later claim resurrects it")
+	}
+	if !ln.Disabled {
+		t.Fatal("placeholder is not marked disabled")
+	}
+}
+
+// TestReconcileLinkDisabledIgnoresEnableForUnknown: an enable for a forward this
+// node has never seen must NOT create a row. Creating one would invent forwards
+// out of ring state.
+func TestReconcileLinkDisabledIgnoresEnableForUnknown(t *testing.T) {
+	st, err := New(filepath.Join(t.TempDir(), "test.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer st.Close()
+	seed(t, st, []string{"other"}, "srv")
+
+	if err := st.ReconcileLinkDisabled("unknown", "srv", 1234, false); err != nil {
+		t.Fatal(err)
+	}
+	if _, found, _ := st.LinkByTriple("unknown", "srv", 1234); found {
+		t.Fatal("an enable for an unknown forward must not materialise a row")
+	}
+}

internal/store/store.go

@@ -128,6 +128,11 @@ type Settings struct {
 	// key). Cleared on explicit detach (detachAsStandalone) so a node that
 	// intentionally left does NOT auto-rejoin.
 	ClusterPeers string `json:"clusterPeers,omitempty"`
+	// BalanceMode is the load-balancing mode chosen in the UI: "", "off",
+	// "dry-run" or "on". Empty means "never configured here" so the environment
+	// variable stays the bootstrap default; a non-empty value wins over it,
+	// otherwise flipping the switch in the UI would silently revert on restart.
+	BalanceMode string `json:"balanceMode,omitempty"`
 }
 
 // Forward is a rendered link row attached to a remote.
@@ -560,6 +565,30 @@ func (s *Store) GetLink(id int64) (Link, bool) {
 	return l, true
 }
 
+// LinkByTriple looks a link up by its natural key (local, remote, remotePort).
+//
+// Prefer this over GetLink(id) whenever the caller only knows the forward's
+// identity: ReplaceLinks() rewrites the whole table with DELETE + re-INSERT, so
+// every row gets a fresh autoincrement id. Any id captured before such a write
+// (e.g. a Link carried inside a ring token) is stale by definition and will
+// either miss or — worse — match a different forward. The natural key is
+// stable across those rewrites.
+//
+// Returns (link, found). A missing row is (Link{}, false) and is NOT an error:
+// callers use that to mean "no persisted opinion yet".
+func (s *Store) LinkByTriple(local, remote string, port int) (Link, bool, error) {
+	var l Link
+	err := s.db.QueryRow("SELECT id, local, remote, remote_port, offset_x, offset_y, grp, disabled FROM links WHERE local = ? AND remote = ? AND remote_port = ?", local, remote, port).
+		Scan(&l.ID, &l.Local, &l.Remote, &l.RemotePort, &l.OffsetX, &l.OffsetY, &l.Group, &l.Disabled)
+	if err != nil {
+		if errors.Is(err, sql.ErrNoRows) {
+			return Link{}, false, nil
+		}
+		return Link{}, false, err
+	}
+	return l, true, nil
+}
+
 // DeleteLink removes a single link by id.
 func (s *Store) DeleteLink(id int64) error {
 	_, err := s.db.Exec("DELETE FROM links WHERE id = ?", id)
@@ -639,6 +668,11 @@ func (s *Store) ReplaceLinks(links []Link) error {
 // SetLinkDisabled flips the disabled flag of a forward identified by its
 // (local, remote, remotePort) natural key. This is the persistence half of the
 // forwards-page start/stop toggle; the caller also drives the worker/ring side.
+//
+// Kept as a targeted UPDATE rather than a ReplaceLinks rewrite on purpose:
+// ReplaceLinks deletes and reinserts every row, handing out fresh autoincrement
+// ids and invalidating any Link a caller captured earlier (they travel inside
+// ring tokens). Flipping one flag must not perturb other rows' identity.
 func (s *Store) SetLinkDisabled(local, remote string, port int, disabled bool) error {
 	_, err := s.db.Exec(
 		"UPDATE links SET disabled = ? WHERE local = ? AND remote = ? AND remote_port = ?",
@@ -647,6 +681,52 @@ func (s *Store) SetLinkDisabled(local, remote string, port int, disabled bool) e
 	return err
 }
 
+// ReconcileLinkDisabled applies a cluster-wide view of one forward's disabled
+// flag into the local store, creating a placeholder row when this node has none
+// yet.
+//
+// This is the receive half of disabled-flag propagation. The stop decision is
+// made on whichever node served the request, then rides the token ring in the
+// topology entry; every other member calls this on adoption so its own
+// links table agrees. Without it the flag lived only on the node that handled
+// the request, and the node actually OWNS the forward — usually a different
+// machine — still believed the forward was enabled and re-spawned its worker.
+//
+// A placeholder row is deliberate: a node that has never seen the forward still
+// needs to remember "this is stopped" so a later claim on this node cannot
+// resurrect it. The placeholder carries the same natural key, so a subsequent
+// real claim fills in the rest.
+func (s *Store) ReconcileLinkDisabled(local, remote string, port int, disabled bool) error {
+	cur, found, err := s.LinkByTriple(local, remote, port)
+	if err != nil {
+		return err
+	}
+	if found {
+		if cur.Disabled == disabled {
+			return nil // already agrees; avoid needless writes every token cycle
+		}
+		return s.SetLinkDisabled(local, remote, port, disabled)
+	}
+	if !disabled {
+		// Nothing to remember: an unknown forward with no entry is simply
+		// "not stopped", which is the default the claim path already assumes.
+		return nil
+	}
+	// Need a placeholder, which requires the local/remote foreign keys to exist.
+	if _, ok := s.GetLocal(local); !ok {
+		return nil // cannot materialise a link without its local peer row
+	}
+	if _, ok := s.GetRemote(remote); !ok {
+		return nil
+	}
+	links, err := s.ListLinks()
+	if err != nil {
+		return err
+	}
+	links = append(links, Link{Local: local, Remote: remote, RemotePort: port, Disabled: true})
+	return s.ReplaceLinks(links)
+}
+
 // SetLinkGroup assigns a management group label to a forward identified by its
 // (local, remote, remotePort) natural key. Empty string clears the group
 // (moves the forward to 未分组). This is the persistence half of the
@@ -673,6 +753,13 @@ func (s *Store) Settings() (Settings, error) {
 	if err := json.Unmarshal([]byte(raw), &st); err != nil {
 		return Settings{AutoStartProfiles: true, RestartOnExit: true, RestartIntervalSeconds: 5}, nil
 	}
+	// BalanceMode lives in its own column (see UpdateSettings). Surface it here
+	// too so callers reading Settings get the effective value, but never let it
+	// overwrite the column on a write-back: UpdateSettings only mirrors a
+	// non-empty value, so an unrelated settings PUT cannot clear a UI choice.
+	if mode, ok, err := s.BalanceMode(); err == nil && ok {
+		st.BalanceMode = mode
+	}
 	return st, nil
 }
 
@@ -685,7 +772,36 @@ func (s *Store) UpdateSettings(st Settings) error {
 		"INSERT INTO settings(key, value) VALUES('settings', ?) ON CONFLICT(key) DO UPDATE SET value=excluded.value",
 		string(raw),
 	)
-	return err
+	if err != nil {
+		return err
+	}
+	if mode := st.BalanceMode; mode != "" {
+		// Mirror into its own column for the same reason the other flagged
+		// fields have one: the settings JSON is rewritten wholesale, so a
+		// clustered field that rides that JSON would flip back and forth between
+		// nodes instead of converging.
+		if _, err := s.db.Exec(
+			"INSERT INTO settings(key, value) VALUES('balance_mode', ?) ON CONFLICT(key) DO UPDATE SET value=excluded.value",
+			mode,
+		); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// BalanceMode returns the stored load-balancing mode, and whether one was ever
+// stored (false = never configured here; the environment default applies).
+func (s *Store) BalanceMode() (string, bool, error) {
+	var v string
+	err := s.db.QueryRow("SELECT value FROM settings WHERE key = 'balance_mode'").Scan(&v)
+	if errors.Is(err, sql.ErrNoRows) {
+		return "", false, nil
+	}
+	if err != nil {
+		return "", false, err
+	}
+	return v, v != "", nil
 }
 
 // SetNodeKey persists just the cluster admission key, preserving all other

internal/version/version.go

@@ -2,11 +2,11 @@
 //
 // Value is overridable at build time so release artifacts carry the real tag:
 //
-//	go build -ldflags="-X webui4frpc/internal/version.Version=0.1.1" ./cmd/webui4frpc
+//	go build -ldflags="-X webui4frpc/internal/version.Version=0.1.4" ./cmd/webui4frpc
 //
 // Kept in its own package because both cmd/ and internal/httpapi report it
 // (main can't be imported, and httpapi must not depend on cmd).
 package version
 
 // Version is the running build's semantic version, without a leading "v".
-var Version = "0.1.1"
+var Version = "0.1.4"

packaging/rpm/webui4frpc.spec

@@ -1,5 +1,5 @@
 Name:           webui4frpc
-Version:        %{?_w4f_version}%{!?_w4f_version:0.1.1}
+Version:        %{?_w4f_version}%{!?_w4f_version:0.1.4}
 Release:        1%{?dist}
 Summary:        Visual frpc controller with token-ring cluster
 License:        MIT
@@ -67,6 +67,21 @@ fi
 %doc %{_docdir}/webui4frpc/README.md
 
 %changelog
+* Mon Oct 06 2026 JianFeeeee <jianfeeeee@gitcode> - 0.1.4-1
+- 负载均衡：改为每节点自主释放（持有数超过平摊份额且 >=2 时自撤销一条丢回环，由最闲节点认领），不再由 leader 统一调度
+- 负载均衡：新增自限上界与下限（撤到平摊值即停、低于 2 条不撤），10 分钟冷却，默认不动作（dry-run）
+- 集群页：新增负载均衡视图与每节点模式开关（关闭/观察/启用），模式持久化
+- 新增只读接口 GET /cluster/balance 与写接口 POST /cluster/balance/mode
+* Sun Oct 05 2026 JianFeeeee <jianfeeeee@gitcode> - 0.1.3-1
+- 窄屏布局：补 box-sizing 重置、接管画布 fit（缩放下限 0.65）、弹窗宽度钳制、统一滚动容器（修复手机端整页无法滑动）
+- 集群：停用/启动改为经令牌环跨节点传播，撤销不再静默失效
+- 集群：修复滚动重启导致的分裂脑、令牌环冻结、非 leader 不重建 worker
+- 集群：令牌轮转日志降级为 debug（355 → 1 行/分钟）
+* Wed Sep 03 2026 JianFeeeee <jianfeeeee@gitcode> - 0.1.2-1
+- 修复窄屏底部导航栏错位到顶部：顶栏的 backdrop-filter 会为
+  position:fixed 后代创建包含块，使 tab bar 的 bottom:0 相对顶栏而
+  非视口定位。窄屏下关掉顶栏模糊，改用近实色底。
+
 * Wed Sep 02 2026 JianFeeeee <jianfeeeee@gitcode> - 0.1.1-1
 - 修复令牌环冻结：心跳到达时复活被误判 offline 的前驱；
   全环无可达后继时保留 inflight，让 leader 超时重发令牌。

packaging/windows/README.md

@@ -79,7 +79,7 @@ cd web && npm run build && rm -rf internal/httpapi/dist && cp -r web/dist intern
 ### Linux (Debian / Ubuntu)
 
 ```bash
-sudo dpkg -i webui4frpc-0.1.1-linux-amd64.deb   # 或 linux-arm64.deb
+sudo dpkg -i webui4frpc-0.1.4-linux-amd64.deb   # 或 linux-arm64.deb
 sudo nano /etc/webui4frpc/webui4frpc.env        # 改掉默认密码
 sudo systemctl restart webui4frpc
 ```
@@ -87,7 +87,7 @@ sudo systemctl restart webui4frpc
 ### Linux (RHEL / Fedora / openEuler / 默认 / 龙蜥)
 
 ```bash
-sudo rpm -i webui4frpc-0.1.1-linux-x86_64.rpm
+sudo rpm -i webui4frpc-0.1.4-linux-x86_64.rpm
 sudo vi /etc/webui4frpc/webui4frpc.env
 sudo systemctl enable --now webui4frpc
 ```
@@ -97,7 +97,7 @@ sudo systemctl enable --now webui4frpc
 
 ### Windows
 
-双击 `webui4frpc-0.1.1-windows-amd64-setup.exe`（ARM 设备用 `windows-arm64-setup.exe`）。
+双击 `webui4frpc-0.1.4-windows-amd64-setup.exe`（ARM 设备用 `windows-arm64-setup.exe`）。
 安装向导会询问监听地址与管理员凭据，完成后：
 
 - 程序装到 `C:\Program Files\webui4frpc`
@@ -107,8 +107,8 @@ sudo systemctl enable --now webui4frpc
 ### macOS
 
 ```bash
-sudo installer -pkg webui4frpc-0.1.1-darwin-arm64.pkg -target /   # Apple Silicon
-sudo installer -pkg webui4frpc-0.1.1-darwin-amd64.pkg -target /   # Intel
+sudo installer -pkg webui4frpc-0.1.4-darwin-arm64.pkg -target /   # Apple Silicon
+sudo installer -pkg webui4frpc-0.1.4-darwin-amd64.pkg -target /   # Intel
 ```
 
 或直接双击 `.pkg` 走图形安装器。安装后 launchd 开机自启，配置在
@@ -126,7 +126,7 @@ sudo launchctl kickstart -k system/com.jianfeeeee.webui4frpc
 不想装服务只想跑一下，用对应的 `.tar.gz` / `.zip`，解压即用：
 
 ```bash
-tar xzf webui4frpc-0.1.1-linux-amd64.tar.gz
+tar xzf webui4frpc-0.1.4-linux-amd64.tar.gz
 ./webui4frpc -addr 127.0.0.1:7500 -user admin -password 你的密码 -workdir ./data
 ```
 

scripts/build_installers.sh

@@ -12,7 +12,7 @@
 # 在 release.sh 流程内被调用；依赖 dist/artifacts/<version>/bin/<os>-<arch>/ 下的二进制。
 set -euo pipefail
 
-VERSION="${1:-0.1.1}"
+VERSION="${1:-0.1.4}"
 VERSION="${VERSION#v}"
 ROOT="$(cd "$(dirname "$0")/.." && pwd)"
 OUT="$ROOT/dist/artifacts/$VERSION"

scripts/probe-with-virtual-frps.sh

@@ -0,0 +1,137 @@
+#!/usr/bin/env bash
+# 用隔离的虚拟 frps 探测 webui4frpc 行为。
+#
+# 全部端口与现网隔离（现网：7000/4434/7500/18082），不修改任何生产配置。
+# 用法：scripts/probe-with-virtual-frps.sh
+set -euo pipefail
+
+LAB="${LAB:-/tmp/frplab}"
+FRP_VER="${FRP_VER:-0.61.1}"
+FRP_TGZ="frp_${FRP_VER}_linux_amd64"
+UI_ADDR="${UI_ADDR:-127.0.0.1:17500}"
+UI_USER="${UI_USER:-labadmin}"
+UI_PASS="${UI_PASS:-labpass123}"
+LAB_TOKEN="${LAB_TOKEN:-lab-token-$(date +%Y%m%d)}"
+
+# 现网 frpc 二进制（唯一允许从生产读取的东西，只读）
+SRC_FRPC="${SRC_FRPC:-/opt/webui4frpc/data/bin/frpc-${FRP_VER}/frpc}"
+BACKEND_PORT=17090
+TCP_PORT=17091
+UI_FORWARD_PORT=17092
+
+cleanup() {
+  echo "==> 清理"
+  pkill -f "webui4frpc -addr ${UI_ADDR}" 2>/dev/null || true
+  pkill -f "frplab/.*frpc" 2>/dev/null || true
+  [[ -n "${FRPS_PID:-}" ]] && kill "${FRPS_PID}" 2>/dev/null || true
+  [[ -n "${BACKEND_PID:-}" ]] && kill "${BACKEND_PID}" 2>/dev/null || true
+}
+trap cleanup EXIT
+
+mkdir -p "${LAB}"/{frps,probe,logs,ui/data}
+
+# ---------- 1. frps（虚拟服务端） ----------
+if [[ ! -x "${LAB}/frpsd" ]]; then
+  echo "==> 下载 frps ${FRP_VER}"
+  curl -sL -o "${LAB}/frps-bin" \
+    "https://github.com/fatedier/frp/releases/download/v${FRP_VER}/${FRP_TGZ}.tar.gz"
+  tar xzf "${LAB}/frps-bin" -C "${LAB}"
+  cp "${LAB}/${FRP_TGZ}/frps" "${LAB}/frpsd"
+  chmod +x "${LAB}/frpsd"
+fi
+
+cat > "${LAB}/frps/frps.toml" <<EOF
+bindPort = 17000
+vhostHTTPPort = 17080
+vhostHTTPsPort = 17443
+auth.token = "${LAB_TOKEN}"
+log.to = "${LAB}/logs/frps.log"
+log.level = "debug"
+EOF
+
+echo "==> 启动虚拟 frps (17000/17080/17443)"
+"${LAB}/frpsd" -c "${LAB}/frps/frps.toml" > "${LAB}/logs/frps-stdout.log" 2>&1 &
+FRPS_PID=$!
+sleep 3
+ss -ltn | grep -E ':17000|:17443' || { echo "frps 未监听"; exit 1; }
+
+# ---------- 2. 后端服务 ----------
+echo "==> 启动后端 http://127.0.0.1:${BACKEND_PORT}"
+( cd "${LAB}" && python3 -m http.server "${BACKEND_PORT}" --bind 127.0.0.1 > logs/backend.log 2>&1 & )
+BACKEND_PID=$!
+sleep 1
+
+# ---------- 3. 裸 frpc 验证 frp 协议本身 ----------
+cat > "${LAB}/probe/frpc-probe.json" <<EOF
+{
+  "serverAddr": "127.0.0.1",
+  "serverPort": 17000,
+  "loginFailExit": false,
+  "auth": {"method": "token", "token": "${LAB_TOKEN}"},
+  "proxies": [
+    {"name":"lab-tcp","type":"tcp","localIP":"127.0.0.1","localPort":${BACKEND_PORT},"remotePort":${TCP_PORT}}
+  ]
+}
+EOF
+cp "${SRC_FRPC}" "${LAB}/frpc" && chmod +x "${LAB}/frpc"
+echo "==> 裸 frpc 连接（验证 frp 协议通路）"
+"${LAB}/frpc" -c "${LAB}/probe/frpc-probe.json" > "${LAB}/logs/frpc.log" 2>&1 &
+sleep 4
+grep -q 'login to server success' "${LAB}/logs/frpc.log" \
+  && echo "    frpc 登录成功" || { echo "    frpc 登录失败"; tail -5 "${LAB}/logs/frpc.log"; }
+printf "    tcp ${TCP_PORT} -> "; curl -s -o /dev/null -w "%{http_code}\n" "http://127.0.0.1:${TCP_PORT}/"
+pkill -f 'frplab/frpc' 2>/dev/null || true
+
+# ---------- 4. webui4frpc 接管 ----------
+echo "==> 启动 webui4frpc ${UI_ADDR}（独立 workdir）"
+mkdir -p "${LAB}/ui/data/bin/frpc-${FRP_VER}"
+cp "${SRC_FRPC}" "${LAB}/ui/data/bin/frpc-${FRP_VER}/frpc"
+chmod +x "${LAB}/ui/data/bin/frpc-${FRP_VER}/frpc"
+( setsid /opt/webui4frpc/webui4frpc -addr "${UI_ADDR}" \
+    -user "${UI_USER}" -password "${UI_PASS}" \
+    -workdir "${LAB}/ui/data" > "${LAB}/logs/ui.log" 2>&1 < /dev/null & )
+sleep 5
+ss -ltn | grep -q "${UI_ADDR##*:}" || { echo "webui4frpc 未监听"; exit 1; }
+
+AUTH="${UI_USER}:${UI_PASS}"
+
+echo "==> 注入画布（注意：loopback 会被自动改写为 LAN 地址）"
+curl -s -u "${AUTH}" -X PUT "http://${UI_ADDR}/api/manager/canvas" \
+  -H 'Content-Type: application/json' \
+  -d "{\"remotes\":[{\"name\":\"lab-frps\",\"ip\":\"127.0.0.1\",\"port\":17000,\"token\":\"${LAB_TOKEN}\",\"enabled\":true}],
+       \"locals\":[{\"name\":\"lab-backend\",\"ip\":\"127.0.0.1\",\"port\":${BACKEND_PORT}}],
+       \"links\":[{\"local\":\"lab-backend\",\"remote\":\"lab-frps\",\"remotePort\":${UI_FORWARD_PORT}}]}" \
+  | python3 -m json.tool
+
+echo "==> binaryPath 必须先写入 settings，否则 worker 静默不启动"
+python3 - "${LAB}/ui/data/manager.db" "${LAB}/ui/data/bin/frpc-${FRP_VER}/frpc" <<'PY'
+import json, sqlite3, sys
+db, binary = sys.argv[1], sys.argv[2]
+c = sqlite3.connect(db)
+raw = c.execute("select value from settings where key='settings'").fetchone()[0]
+d = json.loads(raw); d["binaryPath"] = binary
+c.execute("update settings set value=? where key='settings'", (json.dumps(d),))
+c.commit()
+print("    binaryPath 已写入:", binary)
+PY
+
+echo "==> 启动 forward（非 localOnly 需 owner claim，观察 status）"
+curl -s -u "${AUTH}" -X POST "http://${UI_ADDR}/api/manager/forwards/start" \
+  -H 'Content-Type: application/json' \
+  -d "{\"local\":\"lab-backend\",\"remote\":\"lab-frps\",\"remotePort\":${UI_FORWARD_PORT}}"
+echo
+sleep 5
+
+echo "==> 观察结果"
+curl -s -u "${AUTH}" "http://${UI_ADDR}/api/manager/status" | python3 -c "
+import json,sys
+d=json.load(sys.stdin)
+print('  forwards:', [(f['local'], f['active']) for f in d.get('forwards',[])])
+print('  profiles:', [(p['name'], p['process']['state'], p.get('connState')) for p in d.get('profiles',[])])
+"
+echo "  binaryPath: $(curl -s -u "${AUTH}" "http://${UI_ADDR}/api/manager/binary/status")"
+echo "  configs/:   $(ls "${LAB}/ui/data/configs/" 2>/dev/null | tr '\n' ' ')"
+echo "  workers:    $(pgrep -af "${LAB}/ui/data" | wc -l) 个"
+echo
+echo "日志： ${LAB}/logs/"
+echo "复现结论见 docs/how-it-works.md"

scripts/release.sh

@@ -3,7 +3,7 @@
 # webui4frpc 多平台发布脚本
 #
 # 用法:
-#   ./scripts/release.sh [version]        # 默认 version=0.1.1
+#   ./scripts/release.sh [version]        # 默认 version=0.1.4
 #
 # 行为:
 #   1. 构建前端 (web/dist) 并同步到 internal/httpapi/dist (go:embed 用)
@@ -19,7 +19,7 @@
 #   NO_BUILD_FRONTEND  设为 1 跳过前端构建 (用现有 dist)
 set -euo pipefail
 
-VERSION="${1:-0.1.1}"
+VERSION="${1:-0.1.4}"
 VERSION="${VERSION#v}"   # 去掉可能的前导 v
 ROOT="$(cd "$(dirname "$0")/.." && pwd)"
 OUT="$ROOT/dist/artifacts/$VERSION"

scripts/upload_assets.py

@@ -2,7 +2,8 @@
 """上传 release 资产到 gitcode（两步：取签名 URL → PUT 到 OBS）。
 
 用法: upload_assets.py <version> <token> [file...]
-不传 file 时上传该版本目录下所有 tar.gz/zip 与 SHA256SUMS。
+不传 file 时上传该版本目录下全部发布产物：
+  tar.gz / zip / deb / rpm / pkg / -setup.exe / SHA256SUMS
 """
 import json
 import os
@@ -38,6 +39,19 @@ def put_file(url: str, headers: dict, path: str) -> tuple[int, str]:
         return e.code, e.read().decode("utf-8", "replace")[:300]
 
 
+# 发布产物后缀：免安装包 + 原生安装包。
+# 注意 -setup.exe 而不是裸 .exe，避开 bin/ 里的裸二进制。
+ARTIFACT_SUFFIXES = (
+    ".tar.gz", ".zip",              # 免安装
+    ".deb", ".rpm", ".pkg",         # linux / macOS 安装包
+    "-setup.exe",                   # windows 安装器
+)
+
+
+def is_artifact(name: str) -> bool:
+    return name == "SHA256SUMS" or name.endswith(ARTIFACT_SUFFIXES)
+
+
 def main() -> int:
     if len(sys.argv) < 3:
         print(__doc__)
@@ -47,10 +61,7 @@ def main() -> int:
         os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
         "dist", "artifacts", version,
     )
-    files = sys.argv[3:] or sorted(
-        f for f in os.listdir(outdir)
-        if f.endswith((".tar.gz", ".zip")) or f == "SHA256SUMS"
-    )
+    files = sys.argv[3:] or sorted(f for f in os.listdir(outdir) if is_artifact(f))
     failed = 0
     for name in files:
         path = os.path.join(outdir, name)

web/src/App.vue

@@ -384,16 +384,37 @@ const currentLabel = computed(() => nav.value.find((n) => n.key === view.value)?
     align-items: center;
     gap: 10px;
     padding: 10px 12px;
-    overflow: visible;
+    overflow: hidden;
+    /* 兵家必争：任何子项算错都不得弄出横向滚动条 */
+    overflow-x: hidden;
     border-right: none;
     border-bottom: 1px solid var(--w4f-line);
     padding-top: max(10px, env(safe-area-inset-top));
+    /* 关键：backdrop-filter 会为 position:fixed 后代创建【包含块】
+       （与 transform/filter 同理）。若保留它，下面 .sb-nav 的
+       bottom:0 就是相对这条顶栏定位而非视口，底部 tab bar 会贴在
+       顶栏下沿、看起来固定在屏幕顶部。因此窄屏必须关掉顶栏的玻璃
+       模糊，改用接近实色的底，模糊效果留在 .sb-nav 自己身上
+       （元素自身的 backdrop-filter 不影响它自己的定位）。 */
+    backdrop-filter: none;
+    -webkit-backdrop-filter: none;
+    background: color-mix(in srgb, var(--w4f-card-solid) 92%, transparent);
   }
   .sb-brand {
     padding: 0;
     gap: 8px;
     flex: 1 1 auto;
     min-width: 0;
+    overflow: hidden;
+  }
+  /* 品牌文字必须可收缩：“webui4frpc”是不可断行单词，不给 min-width:0
+     + overflow:hidden 的话它的 flex 自动最小尺寸等于 min-content（≨85px），
+     顶栏就拿不出空间给右侧身份区，溢出为横向滚动。 */
+  .sb-brand-text { min-width: 0; overflow: hidden; }
+  .sb-brand-text b {
+    overflow: hidden;
+    text-overflow: ellipsis;
+    white-space: nowrap;
   }
   .sb-logo { width: 30px; height: 30px; border-radius: 10px; }
   .sb-brand-text b { font-size: 14.5px; }
@@ -425,7 +446,10 @@ const currentLabel = computed(() => nav.value.find((n) => n.key === view.value)?
     gap: 2px;
     padding: 5px 2px;
     border-radius: 11px;
-    font-size: 10.5px;
+    /* Keep the tab label at the theme's 11px floor; the 10.5px here (and 9.5px
+     * at ≤380px) read as noise on a phone. "账号与密钥" is still handled by the
+     * .label ellipsis (width:100% + overflow:hidden). */
+    font-size: 11px;
     font-weight: 600;
     text-align: center;
   }
@@ -442,23 +466,30 @@ const currentLabel = computed(() => nav.value.find((n) => n.key === view.value)?
     box-shadow: none;
   }
 
-  /* 身份与主题折进顶条右侧 */
+  /* 身份与主题折进顶条右侧。
+     不再用 vw 限宽：vw 含垂直滚动条宽度，会比实际可用内容宽大；而且
+     max-width 只封顶盒子，内部 flex:0 0 auto 的徒章/按钮仍会撑出去。
+     改为全链路可收缩（min-width:0 + overflow:hidden），由 flex 自行分配。 */
   .sb-foot {
     margin-top: 0;
     padding-top: 0;
     display: flex;
     align-items: center;
     gap: 8px;
-    flex: 0 0 auto;
+    flex: 0 1 auto;
+    min-width: 0;
   }
   .sb-identity {
     margin-bottom: 0;
     padding: 5px 8px;
     font-size: 11.5px;
-    max-width: 42vw;
+    flex: 0 1 auto;
+    min-width: 0;
+    max-width: none;
+    overflow: hidden;
   }
-  .sb-identity .id-name { max-width: 14vw; }
-  .sb-theme { margin-bottom: 0; }
+  .sb-identity .id-name { min-width: 0; max-width: none; }
+  .sb-theme { margin-bottom: 0; flex: 0 0 auto; }
   .sb-theme .sw { width: 22px; height: 22px; border-radius: 7px; }
 
   .main { height: auto; flex: 1; min-height: 0; }
@@ -470,12 +501,15 @@ const currentLabel = computed(() => nav.value.find((n) => n.key === view.value)?
   }
 }
 
-/* 极窄（≤ 380px，iPhone SE / 小屏安卓）：进一步压缩顶条 */
+/* 极窄（≤ 380px，iPhone SE / 小屏安卓）：进一步压缩顶条。
+   角色徒章是 flex:0 0 auto 不可收缩的（“超级管理员”≨70px），极窄下隐去，
+   用户名与退出按钮保留。
+   注意：底部 tab 文字不再随之降到 9.5px —— 它在手机上低于可读下限
+   （见 .sb-i 的 11px 注释），长标签交给 .label 的省略号处理。 */
 @media (max-width: 380px) {
   .sb-brand-text b { font-size: 13px; }
-  .sb-identity { max-width: 38vw; padding: 4px 7px; }
+  .sb-identity { padding: 4px 7px; }
   .sb-identity .id-level { display: none; }
-  .sb-i { font-size: 9.5px; }
   .content { padding-left: 10px; padding-right: 10px; }
 }
 </style>

web/src/api.ts

@@ -2,6 +2,7 @@
 import type {
 	ApiKey,
 	ApiKeyCreated,
+	BalanceReport,
 	BinaryStatus,
 	CacheResp,
 	CanvasData,
@@ -159,6 +160,21 @@ export const api = {
 	// M6 token ring snapshot.
 	ring: () => request<RingSnapshot>("/api/manager/cluster/ring"),
 
+	// Load balancer read-only view. Reports per-node scores, whether anything is
+	// over the threshold, and the migration it would perform. Never mutates the
+	// cluster. 404 when the balancer is not wired up on that node.
+	balance: () => request<BalanceReport>("/api/manager/cluster/balance"),
+
+	// setBalanceMode switches load balancing on THIS node (off | dry-run | on).
+	// Per-node by design: each member releases only its own forwards, so there is
+	// no cluster-wide switch to flip. Requires write level.
+	setBalanceMode: (mode: string) =>
+		request<BalanceReport>("/api/manager/cluster/balance/mode", {
+			method: "POST",
+			headers: { "Content-Type": "application/json" },
+			body: JSON.stringify({ mode }),
+		}),
+
 	// M6 cluster lifecycle (创建/加入/退出/移除).
 	// clusterCreate: reseed THIS node as a fresh standalone leader (创建集群).
 	clusterCreate: () =>

web/src/styles/theme.css

@@ -184,6 +184,20 @@
 
 /* ---- blue theme (ocean × frost) ---- */
 
+/* ---- box model reset ----
+ * This app had NO box-sizing reset, so every element using `width: 100%` with
+ * horizontal padding overflowed by exactly its padding under the default
+ * content-box model. The narrow-screen top bar was the visible victim:
+ * `.sidebar { width: 100%; padding: 10px 12px }` rendered 375+24 = 399px wide
+ * on a 375px viewport (344px at 320px), pushing the theme swatches and logout
+ * button off the right edge where `overflow-x: hidden` silently clipped them.
+ * It also made the desktop sidebar 246+28 = 274px wide instead of 246px.
+ * border-box makes width include padding+border, which every layout here
+ * already assumes (Element Plus sets it on its own components already). */
+*,
+*::before,
+*::after { box-sizing: border-box; }
+
 html,
 body,
 #app {
@@ -346,4 +360,14 @@ a { color: var(--w4f-primary-h); }
 /* ---- Element Plus dialog tweaks to read as glass ---- */
 .el-overlay { backdrop-filter: blur(3px); }
 .el-dialog { border-radius: var(--w4f-radius) !important; box-shadow: var(--w4f-sh-lg) !important; }
+
+/* Dialogs pass a fixed px width (width="420px" / "440px"), which Element Plus
+ * writes into an INLINE `--el-dialog-width` custom property. On phones that is
+ * wider than the viewport (420px on a 375px screen), so the dialog's right edge
+ * — including the header close button — starts off-screen and is only reachable
+ * by scrolling the overlay horizontally. Clamp it to the viewport; !important
+ * is required to beat the inline custom property. */
+@media (max-width: 720px) {
+  .el-dialog { --el-dialog-width: calc(100vw - 24px) !important; }
+}
 .el-message { border-radius: var(--w4f-radius-sm) !important; box-shadow: var(--w4f-sh-lg) !important; }

web/src/types.ts

@@ -9,332 +9,382 @@ export type CanvasRemote = Remote;
 export type CanvasLink = Link;
 
 export interface Local {
-  name: string;
-  ip: string;
-  port: number;
-  protocol: string; // tcp | udp | http | https
-  // localOnly: forward stays on this node (loopback kept), no cluster.
-  localOnly?: boolean;
+	name: string;
+	ip: string;
+	port: number;
+	protocol: string; // tcp | udp | http | https
+	// localOnly: forward stays on this node (loopback kept), no cluster.
+	localOnly?: boolean;
 
-  // M1 advanced transport knobs (optional; empty/zero = frpc defaults)
-  useEncryption?: boolean;
-  useCompression?: boolean;
-  bandwidthLimit?: string;
-  poolCount?: number;
-  metadatas?: Record<string, string>;
-  annotations?: Record<string, string>;
+	// M1 advanced transport knobs (optional; empty/zero = frpc defaults)
+	useEncryption?: boolean;
+	useCompression?: boolean;
+	bandwidthLimit?: string;
+	poolCount?: number;
+	metadatas?: Record<string, string>;
+	annotations?: Record<string, string>;
 
-  // M2 HTTP/HTTPS routing (http/https only)
-  customDomains?: string; // comma-separated
-  subdomain?: string;
-  locations?: string[];
-  hostHeaderRewrite?: string;
-  httpHeaders?: Record<string, string>;
-  basicAuthUser?: string;
-  basicAuthPassword?: string;
+	// M2 HTTP/HTTPS routing (http/https only)
+	customDomains?: string; // comma-separated
+	subdomain?: string;
+	locations?: string[];
+	hostHeaderRewrite?: string;
+	httpHeaders?: Record<string, string>;
+	basicAuthUser?: string;
+	basicAuthPassword?: string;
 
-  // M3 load balancing group + health check
-  lbGroup?: string;
-  lbGroupKey?: string;
-  healthCheckType?: string; // "" | tcp | http
-  healthCheckPath?: string;
-  healthCheckTimeout?: number;
-  healthCheckMaxFailed?: number;
-  healthCheckInterval?: number;
+	// M3 load balancing group + health check
+	lbGroup?: string;
+	lbGroupKey?: string;
+	healthCheckType?: string; // "" | tcp | http
+	healthCheckPath?: string;
+	healthCheckTimeout?: number;
+	healthCheckMaxFailed?: number;
+	healthCheckInterval?: number;
 }
 
 export interface Remote {
-  name: string;
-  ip: string;
-  port: number; // frps connect port
-  token?: string;
-  url?: string;
-  enabled: boolean;
+	name: string;
+	ip: string;
+	port: number; // frps connect port
+	token?: string;
+	url?: string;
+	enabled: boolean;
 
-  // M1 transport section
-  transportProtocol?: string; // tcp | quic | kcp | websocket
-  transportTls?: boolean;
-  transportPool?: number;
-  transportTlsServerName?: string;
+	// M1 transport section
+	transportProtocol?: string; // tcp | quic | kcp | websocket
+	transportTls?: boolean;
+	transportPool?: number;
+	transportTlsServerName?: string;
 
-  // M2: frps vhostHTTPPort
-  vhostHttpPort?: number;
+	// M2: frps vhostHTTPPort
+	vhostHttpPort?: number;
 
-  // M3: local frpc admin API (webServer)
-  adminAddr?: string;
-  adminPort?: number;
-  adminUser?: string;
-  adminPassword?: string;
+	// M3: local frpc admin API (webServer)
+	adminAddr?: string;
+	adminPort?: number;
+	adminUser?: string;
+	adminPassword?: string;
 }
 
 export interface Link {
-  id?: number;
-  local: string;
-  remote: string;
-  remotePort: number;
-  offsetX?: number;
-  offsetY?: number;
-  // group: management label for one-click group start/stop on the forwards
-  // page (unrelated to frps load-balancing lbGroup on Local).
-  group?: string;
-  // disabled: per-forward stop flag. A disabled forward is omitted from the
-  // frpc config (local) / left out of the ring topology (cluster), so stop is
-  // per-card and survives canvas re-saves.
-  disabled?: boolean;
+	id?: number;
+	local: string;
+	remote: string;
+	remotePort: number;
+	offsetX?: number;
+	offsetY?: number;
+	// group: management label for one-click group start/stop on the forwards
+	// page (unrelated to frps load-balancing lbGroup on Local).
+	group?: string;
+	// disabled: per-forward stop flag. A disabled forward is omitted from the
+	// frpc config (local) / left out of the ring topology (cluster), so stop is
+	// per-card and survives canvas re-saves.
+	disabled?: boolean;
 }
 
 export interface CanvasData {
-  locals: Local[];
-  remotes: Remote[];
-  links: Link[];
+	locals: Local[];
+	remotes: Remote[];
+	links: Link[];
 }
 
 export interface Settings {
-  autoStartProfiles: boolean;
-  restartOnExit: boolean;
-  restartIntervalSeconds: number;
-  binaryPath?: string;
+	autoStartProfiles: boolean;
+	restartOnExit: boolean;
+	restartIntervalSeconds: number;
+	binaryPath?: string;
 }
 
 export interface ProcessStatus {
-  state: string;
-  pid?: number;
-  startTime?: number;
-  restartCount: number;
-  exitCode?: number;
-  err?: string;
+	state: string;
+	pid?: number;
+	startTime?: number;
+	restartCount: number;
+	exitCode?: number;
+	err?: string;
 }
 
 export interface ProxyState {
-  name: string;
-  type: string;
-  status: string; // new | wait start | start error | running | check failed | closed
-  err?: string;
-  local_addr?: string;
-  remote_addr?: string;
+	name: string;
+	type: string;
+	status: string; // new | wait start | start error | running | check failed | closed
+	err?: string;
+	local_addr?: string;
+	remote_addr?: string;
 }
 
 export interface StatusProfile {
-  name: string;
-  enabled: boolean;
-  process: ProcessStatus;
-  hasProcess: boolean;
-  forwards: { service: string; remotePort: number; localPort?: number }[];
-  // connState: derived frps connection state (we do not control remote frps)
-  connState?: string; // connected | connecting | failed | not_started | disabled
-  // M3: admin API state
-  adminEnabled?: boolean;
-  proxyStates?: ProxyState[];
-  groupCount?: number;
+	name: string;
+	enabled: boolean;
+	process: ProcessStatus;
+	hasProcess: boolean;
+	forwards: { service: string; remotePort: number; localPort?: number }[];
+	// connState: derived frps connection state (we do not control remote frps)
+	connState?: string; // connected | connecting | failed | not_started | disabled
+	// M3: admin API state
+	adminEnabled?: boolean;
+	proxyStates?: ProxyState[];
+	groupCount?: number;
 }
 
 export interface StatusResp {
-  version: string;
-  workDir: string;
-  settings: Settings;
-  services: Local[];
-  remotes: Remote[];
-  binaryPath: string;
-  profiles: StatusProfile[];
-  localStatus: LocalStatus[];
-  // forwards: the forward-centric list rendered as cards on the status page.
-  forwards: ForwardStatus[];
-  // selfId: this node's ring ID (== SelfAddr), to tell whether a cluster
-  // forward's owner is this node (worker runs here) or a peer.
-  selfId: string;
+	version: string;
+	workDir: string;
+	settings: Settings;
+	services: Local[];
+	remotes: Remote[];
+	binaryPath: string;
+	profiles: StatusProfile[];
+	localStatus: LocalStatus[];
+	// forwards: the forward-centric list rendered as cards on the status page.
+	forwards: ForwardStatus[];
+	// selfId: this node's ring ID (== SelfAddr), to tell whether a cluster
+	// forward's owner is this node (worker runs here) or a peer.
+	selfId: string;
 }
 
 // ForwardStatus is one forward (link) in the forward-centric status view.
 // kind distinguishes 本地转发 (localOnly) from 远程转发 (cluster-distributed).
 export interface ForwardStatus {
-  local: string;
-  remote: string;
-  remotePort: number;
-  localOnly: boolean;
-  kind: "local" | "remote";
-  ownerId?: string; // remote = topology owner; local = selfId
-  active: boolean; // remote = in topology; local = local worker running
-  disabled: boolean; // per-forward stop flag (true = stopped)
-  group?: string;
-  localIp?: string;
-  localPort?: number;
-  localProto?: string;
+	local: string;
+	remote: string;
+	remotePort: number;
+	localOnly: boolean;
+	kind: "local" | "remote";
+	ownerId?: string; // remote = topology owner; local = selfId
+	active: boolean; // remote = in topology; local = local worker running
+	disabled: boolean; // per-forward stop flag (true = stopped)
+	group?: string;
+	localIp?: string;
+	localPort?: number;
+	localProto?: string;
 }
 
 export interface LocalTargetStatus {
-  remote: string;
-  remotePort: number;
-  workerState: string;
+	remote: string;
+	remotePort: number;
+	workerState: string;
 }
 
 export interface LocalStatus {
-  local: Local;
-  targets: LocalTargetStatus[];
+	local: Local;
+	targets: LocalTargetStatus[];
 }
 
 export interface BinaryStatus {
-  binaryPath: string;
-  version?: string;
+	binaryPath: string;
+	version?: string;
 }
 
 export interface InstallResult {
-  path: string;
-  version: string;
+	path: string;
+	version: string;
 }
 
 // M6: cluster node registry + binary cache.
 export interface ClusterNode {
-  addr: string;
-  cache?: string[];
-  version?: string;
+	addr: string;
+	cache?: string[];
+	version?: string;
 }
 
 export interface CacheEntry {
-  version: string;
-  path: string;
-  size: number;
-  modTime: number;
+	version: string;
+	path: string;
+	size: number;
+	modTime: number;
 }
 
 export interface ClusterNodesResp {
-  nodes: ClusterNode[];
+	nodes: ClusterNode[];
 }
 
 export interface CacheResp {
-  cache: CacheEntry[];
-  removed?: string[];
+	cache: CacheEntry[];
+	removed?: string[];
 }
 
 // M6 token-ring snapshot (GET /cluster/ring).
 export interface RingNode {
-  id: string;
-  addr: string;
-  isLeader?: boolean;
-  alive: boolean;
-  load: { memPct: number; netPct: number; forwards?: number };
-  version?: string;
-  cache?: string[];
-  lastSeen?: number;
+	id: string;
+	addr: string;
+	isLeader?: boolean;
+	alive: boolean;
+	load: { memPct: number; netPct: number; forwards?: number };
+	version?: string;
+	cache?: string[];
+	lastSeen?: number;
 }
 
 export interface RingTaskInfo {
-  id: string;
-  revoke?: boolean;
-  removeNode?: string; // node-removal command (target rides the token to self-remove)
-  local: { name: string };
-  remote: { name: string };
-  link: { remotePort?: number };
+	id: string;
+	revoke?: boolean;
+	removeNode?: string; // node-removal command (target rides the token to self-remove)
+	local: { name: string };
+	remote: { name: string };
+	link: { remotePort?: number };
 }
 
 export interface RingTopoEntry {
-  taskId: string;
-  ownerId: string;
-  active: boolean;
-  local: { name: string };
-  remote: { name: string };
-  link: { remotePort?: number };
+	taskId: string;
+	ownerId: string;
+	active: boolean;
+	local: { name: string };
+	remote: { name: string };
+	link: { remotePort?: number };
 }
 
 export interface RingLogEntry {
-  seq: number;
-  node: string;
-  kind: string;
-  at?: number;
-  data?: Record<string, any> | null;
+	seq: number;
+	node: string;
+	kind: string;
+	at?: number;
+	data?: Record<string, any> | null;
 }
 
 export interface RingSnapshot {
-  selfId: string;
-  leaderId: string; // 当前 leader（持首令牌 + 判丢失 + 被上家监控）
-  cycle: number; // 周期号（leader 每收令牌回 +1）
-  lastSync: number; // 上次同步时间 (unix sec)
-  roundDelayMs?: number;
-  nodes: RingNode[];
-  pending: RingTaskInfo[];
-  topology: RingTopoEntry[];
-  log: RingLogEntry[];
-  // nodeKey: this node's cluster admission key. Newcomers must present it to
-  // join via this node (handleClusterJoin verifies). Shown on the cluster page
-  // for the operator to copy; omitted for non-cluster (standalone) nodes.
-  nodeKey?: string;
+	selfId: string;
+	leaderId: string; // 当前 leader（持首令牌 + 判丢失 + 被上家监控）
+	cycle: number; // 周期号（leader 每收令牌回 +1）
+	lastSync: number; // 上次同步时间 (unix sec)
+	roundDelayMs?: number;
+	nodes: RingNode[];
+	pending: RingTaskInfo[];
+	topology: RingTopoEntry[];
+	log: RingLogEntry[];
+	// nodeKey: this node's cluster admission key. Newcomers must present it to
+	// join via this node (handleClusterJoin verifies). Shown on the cluster page
+	// for the operator to copy; omitted for non-cluster (standalone) nodes.
+	nodeKey?: string;
+}
+
+// BalanceReport is one node's own view of rebalancing (GET /cluster/balance).
+//
+// It is per-node by design: every member runs the same rule against its own
+// forwards, and the answer that matters is the one on the node doing (or
+// refusing to do) the releasing. So this reports "my share" first and the peers
+// second, and there is no cluster-wide "would migrate" plan — the ring's normal
+// lowest-load claim decides where a released forward lands.
+export interface BalanceReleasePlan {
+	local: string;
+	remote: string;
+	remotePort: number;
+	owner: string;
+}
+
+export interface BalanceReport {
+	mode: "off" | "dry-run" | "on";
+	modeSource: "env" | "stored"; // env = W4F_BALANCE_MODE default, stored = set here
+	enabled: boolean; // mode !== off
+	dryRun: boolean; // mode !== on
+	canControl: boolean; // this principal may switch it (write level)
+	isMember: boolean;
+	self: string;
+
+	selfForwards: number;
+	share: number; // the even split: ceil(total / alive)
+	totalForwards: number;
+	aliveNodes: number;
+	selfOverShare: boolean;
+
+	// wouldRelease is present whenever this node is over its share — including
+	// while it is deliberately holding back, which is what blockedBy explains.
+	wouldRelease?: BalanceReleasePlan;
+	releasing?: BalanceReleasePlan;
+	blockedBy?: string; // single-node | below-min | balanced | nothing-active | cooldown | in-flight
+
+	cooldownSec: number;
+	lastReleaseAt?: number;
+	nodes: BalanceNodeReport[];
+}
+
+export interface BalanceNodeReport {
+	id: string;
+	alive: boolean;
+	isLeader: boolean;
+	forwards: number;
+	memPct: number;
+	netPct: number;
+	score: number;
 }
 
 // ---- Auth / accounts / API keys (M7) ----
 
-export type AccessLevel = 'read' | 'write' | 'admin' | 'superadmin';
+export type AccessLevel = "read" | "write" | "admin" | "superadmin";
 
 // MeResp is the authenticated principal (GET /me). The frontend uses level to
 // gate the UI: viewer (= auditors) sees read + export controls only.
 export interface MeResp {
-  type: 'user' | 'service' | 'key';
-  name: string;
-  level: AccessLevel;
-  userId?: number;
+	type: "user" | "service" | "key";
+	name: string;
+	level: AccessLevel;
+	userId?: number;
 }
 
 export interface User {
-  id: number;
-  username: string;
-  role: 'admin' | 'viewer' | 'superadmin';
-  enabled: boolean;
-  system?: boolean; // flag-synced built-in account (UI read-only)
-  createdAt: number;
-  lastLoginAt: number;
+	id: number;
+	username: string;
+	role: "admin" | "viewer" | "superadmin";
+	enabled: boolean;
+	system?: boolean; // flag-synced built-in account (UI read-only)
+	createdAt: number;
+	lastLoginAt: number;
 }
 
 export interface ApiKey {
-  id: number;
-  userId: number;
-  prefix: string;
-  label: string;
-  scope: AccessLevel;
-  createdAt: number;
-  lastUsedAt: number;
-  expiresAt?: number;
+	id: number;
+	userId: number;
+	prefix: string;
+	label: string;
+	scope: AccessLevel;
+	createdAt: number;
+	lastUsedAt: number;
+	expiresAt?: number;
 }
 
 // ApiKeyCreated is the one-time response from POST /apikeys: the plaintext key
 // is shown exactly once and never retrievable again.
 export interface ApiKeyCreated {
-  id: number;
-  userId: number;
-  label: string;
-  scope: AccessLevel;
-  prefix: string;
-  createdAt: number;
-  key: string;
+	id: number;
+	userId: number;
+	label: string;
+	scope: AccessLevel;
+	prefix: string;
+	createdAt: number;
+	key: string;
 }
 
 export interface CanvasExportEnvelope {
-  _type: 'webui4frpc-canvas';
-  version: number;
-  exportedAt: number;
-  exporter: string;
-  canvas: CanvasData;
+	_type: "webui4frpc-canvas";
+	version: number;
+	exportedAt: number;
+	exporter: string;
+	canvas: CanvasData;
 }
 
 // Worker-log bundle (GET /cluster/logs/export): one entry per alive ring node.
 export interface WorkerLogEntry {
-  name: string;
-  remote: string;
-  state: string;
-  lines: string;
+	name: string;
+	remote: string;
+	state: string;
+	lines: string;
 }
 export interface NodeLogsResp {
-  node: string;
-  workers: WorkerLogEntry[];
+	node: string;
+	workers: WorkerLogEntry[];
 }
 export interface ClusterWorkerLogNode {
-  id: string;
-  addr: string;
-  ok: boolean;
-  error?: string;
-  logs?: NodeLogsResp;
+	id: string;
+	addr: string;
+	ok: boolean;
+	error?: string;
+	logs?: NodeLogsResp;
 }
 export interface WorkerLogBundle {
-  _type: 'webui4frpc-worker-logs';
-  version: number;
-  exportedAt: number;
-  requester: string;
-  nodes: ClusterWorkerLogNode[];
+	_type: "webui4frpc-worker-logs";
+	version: number;
+	exportedAt: number;
+	requester: string;
+	nodes: ClusterWorkerLogNode[];
 }

web/src/views/CanvasView.vue

@@ -25,16 +25,17 @@
       <VueFlow
         v-model:nodes="nodes"
         v-model:edges="edges"
-        :default-viewport="{ zoom: 0.85 }"
-        :min-zoom="0.3"
+        :default-viewport="{ zoom: FIT_MAX_ZOOM }"
+        :min-zoom="MIN_ZOOM"
         :max-zoom="2"
-        fit-view-on-init
         :delete-key-code="null"
         :nodes-draggable="canWrite"
         :nodes-connectable="canWrite"
         :edges-updatable="false"
         :edges-reconnectable="false"
         class="flow-canvas"
+        @init="onFlowInit"
+        @nodes-initialized="reframe"
         @connect="onConnect"
         @pane-click="deselectAll"
         @edge-click="onEdgeClick"
@@ -118,7 +119,7 @@
 <script setup lang="ts">
 import '@vue-flow/core/dist/style.css'
 import '@vue-flow/core/dist/theme-default.css'
-import { ref, computed, onMounted, onBeforeUnmount } from 'vue'
+import { ref, computed, onMounted, onBeforeUnmount, nextTick } from 'vue'
 import { VueFlow } from '@vue-flow/core'
 import { Background } from '@vue-flow/background'
 import { ElMessage, ElMessageBox } from 'element-plus'
@@ -134,6 +135,103 @@ const nodes = ref<any[]>([])
 const edges = ref<Edge[]>([])
 const loading = ref(false)
 
+// ---- viewport / zoom floor ----
+// The canvas is laid out in absolute graph units (local column at x=60, remote
+// column at x=760, and BOTH columns grow downward on every added forward), so
+// its bounding box is far wider and taller than a phone viewport. Vue Flow's
+// `fit-view-on-init` frames that whole box, which on a phone produced a scale of
+// ~0.31: node cards measure 248px in graph units but rendered 85px wide at an
+// effective font size of ~4.4px — unreadable. Even a 1280px desktop landed at
+// ~0.60 (8.4px) with five forwards stacked.
+//
+// So the canvas is framed by frameCanvas() below, clamped by a zoom floor, and
+// the canvas pans when the content no longer fits. 0.65 renders a 240px local
+// card at ~156px with a legible ~9px font.
+//
+// NOTE: `fit-view-on-init` is deliberately NOT set. It resolves on its own
+// schedule and overwrote the explicit fit performed here (observed: this code
+// measured the unfitted layout, set the viewport, and then the deferred init fit
+// landed last and won). Owning the fit outright is the only way to keep it
+// deterministic.
+//
+// The store arrives via the `init` event. That is deliberately preferred over
+// `useVueFlow()`: this component is the PARENT of <VueFlow>, and calling
+// useVueFlow() in a parent without a shared id can resolve to a different store
+// than the one the rendered flow uses. The event cannot misfire that way.
+const MIN_ZOOM = 0.65
+const FIT_MAX_ZOOM = 1
+type FlowStore = {
+  fitView: (opts?: Record<string, unknown>) => Promise<boolean>
+  getViewport?: () => { x: number; y: number; zoom: number }
+  setViewport?: (
+    t: { x: number; y: number; zoom: number },
+    opts?: Record<string, unknown>,
+  ) => Promise<boolean>
+}
+let flowStore: FlowStore | null = null
+
+// frameCanvas fits the graph, then re-anchors it to the top-left padding edge.
+//
+// Two reasons it is not just fitView():
+//  1. the zoom floor (MIN_ZOOM) — fitView alone shrank 5 stacked forwards to
+//     ~0.31 on a phone, rendering 4.4px text;
+//  2. fitView centres the bounding box, which on a narrow screen puts the seam
+//     between the local and remote columns mid-viewport — the user sees half of
+//     each. Anchoring the top-left node to the padding edge means whole local
+//     cards are visible immediately and the remote column is a pan away.
+//
+// duration:0 on both transitions — we measure node rects in between, and a
+// running animation would hand back mid-flight transforms and skew the anchor.
+const PANE_PAD_RATIO = 0.06
+const frameCanvas = async () => {
+  if (!flowStore) return
+  await flowStore.fitView({
+    padding: 0.12,
+    minZoom: MIN_ZOOM,
+    maxZoom: FIT_MAX_ZOOM,
+    duration: 0,
+  })
+
+  const pane = document.querySelector<HTMLElement>('.vue-flow')
+  const nodeEls = Array.from(document.querySelectorAll<HTMLElement>('.vue-flow__node'))
+  if (!pane || !nodeEls.length) return
+
+  const paneRect = pane.getBoundingClientRect()
+  const rects = nodeEls.map((el) => el.getBoundingClientRect())
+  const dx =
+    paneRect.left + pane.clientWidth * PANE_PAD_RATIO - Math.min(...rects.map((r) => r.left))
+  const dy =
+    paneRect.top + pane.clientHeight * PANE_PAD_RATIO - Math.min(...rects.map((r) => r.top))
+  if (Math.abs(dx) < 1 && Math.abs(dy) < 1) return
+
+  const vp = flowStore.getViewport?.()
+  if (!vp) return
+  await flowStore.setViewport?.({ x: vp.x + dx, y: vp.y + dy, zoom: vp.zoom }, { duration: 0 })
+}
+
+// The flow store and the node list become available at different times: the
+// store at component init, the nodes only once the async canvas fetch resolves
+// (plus a further tick for Vue Flow to measure them). Framing on any single
+// signal would fit an EMPTY canvas, so every path calls reframe() and it no-ops
+// until there is a store and something to frame.
+//
+// `nodes-initialized` also fires whenever a node is added later. Auto-framing on
+// that would yank the viewport out from under whoever is mid-pan, so the initial
+// frame is claimed exactly once; rearranging is the only thing that re-frames on
+// demand (see autoLayout).
+let framedOnce = false
+const reframe = (force = false) => {
+  if (!flowStore || !nodes.value.length) return
+  if (framedOnce && !force) return
+  framedOnce = true
+  nextTick(() => { void frameCanvas() })
+}
+
+const onFlowInit = (store: FlowStore) => {
+  flowStore = store
+  reframe()
+}
+
 // Background dot-pattern color, resolved from the live theme var so the grid
 // recolors on theme switch (vue-flow Background renders pattern-color as an
 // SVG fill attribute, which can't resolve var() directly — read the computed
@@ -667,7 +765,12 @@ const remoteNodes = computed(() =>
   nodes.value.filter((n) => n.type === 'remote'),
 )
 
-const autoLayout = () => layoutColumns(true)
+// Re-frame after rearranging, bounded by the same zoom floor so "自动排列" never
+// shrinks the cards back into unreadable territory on a narrow screen.
+const autoLayout = () => {
+  layoutColumns(true)
+  reframe(true)
+}
 
 const load = async () => {
   loading.value = true
@@ -718,6 +821,9 @@ const load = async () => {
     }
     assignLayers()
     dirty.value = false
+    // The store exists by now but the node list has only just been populated, so
+    // this is the first moment a fit actually has something to frame.
+    reframe(true)
   } finally {
     loading.value = false
   }

web/src/views/ClusterView.vue

@@ -121,6 +121,131 @@
         </div>
       </section>
 
+      <!-- Load balancer (read-only). Shown only when the endpoint responds: a
+           404 means this node has no balancer wired up, and an absent section is
+           the honest signal for "this build doesn't do balancing". -->
+      <section class="section w4f-card bal-section" v-if="bal">
+        <h3 class="section-title">
+          <span class="k-ic">⚖</span>负载均衡
+          <!-- The switch edits the mode ON THIS NODE. Every member decides about
+               its own forwards, so a cluster-wide toggle would be a lie: there is
+               no single action that flips them together. -->
+          <el-select
+            v-if="bal.canControl"
+            v-model="balModeChoice"
+            size="small"
+            class="bal-mode"
+            :disabled="modeBusy"
+            @change="onModeChange"
+          >
+            <el-option label="关闭" value="off" />
+            <el-option label="观察模式（不释放）" value="dry-run" />
+            <el-option label="启用（会释放）" value="on" />
+          </el-select>
+          <span class="w4f-tag" v-else v-text="'只读'" ></span>
+          <span class="bal-src" v-if="bal.modeSource === 'stored'">已保存</span>
+          <span class="bal-src" v-else>来自启动参数</span>
+          <span class="w4f-tag" v-if="!bal.isMember" v-text="'非集群成员'"></span>
+          <span class="w4f-tag" v-else-if="bal.releasing" v-text="'释放进行中'"></span>
+          <span class="w4f-tag w4f-tag-warn" v-else-if="bal.selfOverShare" v-text="'本机超出份额'"></span>
+          <span class="w4f-tag w4f-tag-ok" v-else v-text="'本机在份额内'"></span>
+          <span class="title-actions">
+            <button class="w4f-btn ghost small" :disabled="busy" @click="load">⟳ 刷新</button>
+          </span>
+        </h3>
+
+        <div class="bal-thresholds">
+          <span>
+            规则：本机转发 <b>{{ bal.selfForwards }}</b> 条
+            &gt; 平摊份额 <b>{{ bal.share }}</b>
+            且至少 {{ 2 }} 条时，本机自行释放一条
+          </span>
+          <span>集群共 {{ bal.totalForwards }} 条 / {{ bal.aliveNodes }} 台存活</span>
+          <span>两次释放间隔 ≥ {{ Math.round(bal.cooldownSec / 60) }} 分钟</span>
+          <span v-if="bal.dryRun" class="bal-hint">
+            设 <code>W4F_BALANCE_MODE=on</code> 才会真正释放
+          </span>
+        </div>
+
+        <p class="bal-note">
+          每个节点各自判断自己的转发，超了就自己撤销一条并丢回环上，
+          由环节点按「最闲者优先」重新认领。没有中心调度。
+        </p>
+
+        <!-- Per-node forward counts. Bars are scaled to the busiest node so the
+             shape reads whether the cluster holds 3 forwards or 300; the share
+             marker is what actually decides, not the bar. -->
+        <div class="bal-nodes">
+          <div
+            v-for="n in balNodes"
+            :key="n.id"
+            class="bal-node"
+            :class="{ self: n.id === bal.self, over: n.forwards > bal.share }"
+          >
+            <div class="bn-top">
+              <span class="bn-id">{{ shortId(n.id) }}</span>
+              <span v-if="n.id === bal.self" class="w4f-tag w4f-tag-primary">本机</span>
+              <span v-if="n.isLeader" class="w4f-tag w4f-tag-info">LEADER</span>
+              <span v-if="!n.alive" class="w4f-tag w4f-tag-danger">离线</span>
+              <span v-if="n.alive && n.forwards > bal.share" class="w4f-tag w4f-tag-warn">超出份额</span>
+            </div>
+            <div class="w4f-bar" :class="{ err: n.alive && n.forwards > bal.share }">
+              <i :style="{ width: barWidth(n.forwards) + '%' }" />
+            </div>
+            <div class="bn-stats">
+              <span>转发 <b>{{ n.forwards }}</b></span>
+              <span>内存 <b>{{ n.memPct.toFixed(1) }}%</b></span>
+              <span>网络 <b>{{ n.netPct.toFixed(1) }}%</b></span>
+              <span class="bn-score">份额 <b>{{ bal.share }}</b></span>
+            </div>
+          </div>
+          <div v-if="!balNodes.length" class="empty">暂无节点负载数据</div>
+        </div>
+
+        <!-- What THIS node would do. The answer is per-node by design: every
+             member runs the same rule on its own forwards, so there is no
+             cluster-wide plan and no "the leader will move X to Y". -->
+        <div class="bal-plan">
+          <div v-if="bal.releasing" class="bp-inflight">
+            <span class="w4f-tag w4f-tag-warn">释放进行中</span>
+            <span class="bp-text">
+              本机已释放 {{ bal.releasing.local }}:{{ bal.releasing.remotePort }}，
+              等待环把它认领到别的节点
+            </span>
+          </div>
+          <div v-else-if="bal.wouldRelease && !bal.blockedBy" class="bp-would">
+            <span class="w4f-tag w4f-tag-warn">将释放</span>
+            <span class="bp-text">
+              本机将撤销 {{ bal.wouldRelease.local }}:{{ bal.wouldRelease.remotePort }}
+              并丢回环上，由最闲节点接管
+            </span>
+          </div>
+          <div v-else-if="bal.wouldRelease && bal.blockedBy === 'cooldown'" class="bp-waiting">
+            <span class="w4f-tag w4f-tag-info">冷却中</span>
+            <span class="bp-text">
+              本机超出份额，但距上次释放不足
+              {{ Math.round(bal.cooldownSec / 60) }} 分钟
+              <template v-if="bal.lastReleaseAt">
+                （已过 {{ cooldownAgo }}）
+              </template>
+            </span>
+          </div>
+          <div v-else-if="bal.wouldRelease" class="bp-waiting">
+            <span class="w4f-tag w4f-tag-info">暂不释放</span>
+            <span class="bp-text">本机超出份额，原因：{{ blockedLabel }}</span>
+          </div>
+          <div v-else-if="!bal.isMember" class="bp-ok">
+            <span class="w4f-tag w4f-tag-ok">单节点</span>
+            <span class="bp-text">本机未加入集群，无需均衡</span>
+          </div>
+          <div v-else class="bp-ok">
+            <span class="w4f-tag w4f-tag-ok">已均衡</span>
+            <span class="bp-text">
+              本机持有 {{ bal.selfForwards }} 条，未超平摊份额 {{ bal.share }}，无需释放
+            </span>
+          </div>
+        </div>
+      </section>
       <!-- Active forward topology -->
       <section class="section w4f-card">
         <h3 class="section-title"><span class="k-ic">→</span>活跃转发拓扑</h3>
@@ -199,14 +324,18 @@
 </template>
 
 <script setup lang="ts">
-import { ref, reactive, computed, onMounted, onBeforeUnmount } from 'vue'
+import { ref, reactive, computed, onMounted, onBeforeUnmount, watch } from 'vue'
 import { ElMessage } from 'element-plus'
 import { api, downloadAuditCsv } from '../api'
 import { canWrite } from '../auth'
-import type { RingSnapshot, RingTaskInfo, RingLogEntry } from '../types'
+import type { BalanceReport, RingSnapshot, RingTaskInfo, RingLogEntry } from '../types'
 
 const ring = ref<RingSnapshot | null>(null)
 const busy = ref(false)
+// bal is null when the node has no balancer wired up (endpoint 404s). The
+// section is simply not rendered then, which is the honest signal for "this
+// build does not do balancing" — better than an empty panel implying it does.
+const bal = ref<BalanceReport | null>(null)
 let timer: number | null = null
 let visHandler: (() => void) | null = null
 
@@ -214,8 +343,76 @@ const joinDialog = reactive({ visible: false, addr: '', key: '' })
 
 const load = async () => {
   try { ring.value = await api.ring() } catch { /* keep last */ }
+  // Loaded alongside the ring, but unlike it a failure clears the view: the
+  // balancer report is a decision surface, and showing a stale "已均衡" after the
+  // endpoint starts failing would be worse than showing nothing.
+  try { bal.value = await api.balance() } catch { bal.value = null }
 }
 
+// ---- load balancer mode switch ---- //
+// The select is bound to a local mirror so the UI can show the pending choice
+// while the request is in flight, and snap back if the server rejects it.
+const balModeChoice = ref<string>('dry-run')
+const modeBusy = ref(false)
+watch(
+  () => bal.value?.mode,
+  (m) => { if (m) balModeChoice.value = m },
+  { immediate: true },
+)
+const onModeChange = async (m: string) => {
+  if (!bal.value || modeBusy.value) return
+  modeBusy.value = true
+  try {
+    bal.value = await api.setBalanceMode(m)
+    ElMessage.success(
+      m === 'on' ? '本机已启用负载释放' : m === 'off' ? '本机已关闭负载均衡' : '本机已切到观察模式',
+    )
+  } catch (e) {
+    // Snap back: leaving the select showing a mode the server did not accept
+    // would misrepresent what the node is doing.
+    balModeChoice.value = bal.value.mode
+    ElMessage.error('切换失败：' + (e instanceof Error ? e.message : String(e)))
+  } finally {
+    modeBusy.value = false
+  }
+}
+
+// ---- load balancer view ---- //
+// Each node decides about its OWN forwards, so this page has no cluster-wide
+// plan to show — only what THIS node holds and what it is about to do.
+// Ordered by descending count: the busiest node first is what an operator
+// scanning for a hotspot wants.
+const balNodes = computed(() => [...(bal.value?.nodes ?? [])].sort((a, b) => b.forwards - a.forwards))
+// barWidth scales against the busiest node rather than an absolute, so the shape
+// is readable whether the cluster holds 3 forwards or 300. The bar is context;
+// `share` is what actually decides whether a node is over.
+const barWidth = (forwards: number) => {
+  const max = Math.max(...(bal.value?.nodes ?? []).map((n) => n.forwards), 1)
+  return Math.min(100, Math.round((forwards / max) * 100))
+}
+// blockedLabel renders the server's machine reason as something readable. The
+// reason matters: "over its share but not acting" is otherwise indistinguishable
+// from "broken".
+const blockedLabel = computed(() => {
+  switch (bal.value?.blockedBy) {
+    case 'cooldown': return '冷却期未过'
+    case 'in-flight': return '上一次释放尚未确认'
+    case 'nothing-active': return '本机可释放的转发都已停用'
+    case 'below-min': return '持有量低于下限'
+    case 'single-node': return '单节点环'
+    default: return bal.value?.blockedBy || '未说明'
+  }
+})
+// cooldownAgo says how long ago the last release was, so an operator can see
+// whether the cooldown is about to expire.
+const cooldownAgo = computed(() => {
+  const t = bal.value?.lastReleaseAt ?? 0
+  if (!t) return ''
+  const s = Math.max(0, Math.floor(nowMs.value / 1000) - t)
+  if (s < 60) return s + ' 秒前'
+  return Math.floor(s / 60) + ' 分钟前'
+})
+
 const shortId = (id: string) => (id.length > 22 ? '…' + id.slice(-20) : id)
 
 // copyNodeKey copies this node's admission key to the clipboard so the operator
@@ -260,6 +457,8 @@ const stateDotClass = computed(() => {
 
 // ---- last-sync freshness (ops cares "环还活着吗", not the cycle counter) ----
 const now = ref(Date.now())
+// nowMs mirrors `now` as a plain number for the balance section's minute math.
+const nowMs = computed(() => now.value)
 let tick: number | null = null
 const lastSyncAgo = computed(() => {
   const ls = ring.value?.lastSync ?? 0
@@ -481,7 +680,11 @@ onBeforeUnmount(() => {
 </script>
 
 <style scoped lang="scss">
-.cluster-page { height: 100%; display: flex; flex-direction: column; overflow-y: auto; padding: 4px 0 40px; gap: 14px; }
+/* Let .content (App.vue) remain the single scroll container — see the note in
+ * StatusView: a page-level scroller sized by leftover flex space is what froze
+ * the status page. `.log-list` keeps its own bounded scroll (that one is
+ * deliberate — a log pane should not push the whole page). */
+.cluster-page { display: flex; flex-direction: column; padding: 4px 0 40px; gap: 14px; }
 
 /* hero */
 .hero { display: flex; align-items: center; justify-content: space-between; gap: 16px; flex-wrap: wrap; padding: 16px 20px; }
@@ -634,4 +837,66 @@ onBeforeUnmount(() => {
   .kpis { grid-template-columns: minmax(0, 1fr); }
   .page-title { font-size: 15.5px; }
 }
+
+/* ---- Load balancer section ---- */
+.bal-section .section-title { display: flex; align-items: center; gap: 8px; flex-wrap: wrap; }
+.bal-section .section-title .k-ic { margin-right: 0; }
+
+.bal-thresholds {
+  display: flex; flex-wrap: wrap; gap: 6px 14px;
+  font-size: 11.5px; color: var(--w4f-muted);
+  margin-bottom: 10px;
+}
+.bal-hint { color: var(--w4f-warning); }
+.bal-hint code {
+  font-family: var(--w4f-mono, ui-monospace, monospace);
+  background: var(--w4f-card-2); padding: 1px 4px; border-radius: 3px;
+}
+
+.bal-nodes {
+  display: grid; gap: 8px;
+  grid-template-columns: repeat(auto-fit, minmax(200px, 1fr));
+}
+.bal-node {
+  border: 1px solid var(--w4f-line);
+  border-radius: 8px;
+  padding: 9px 11px;
+  background: var(--w4f-card-2);
+  min-width: 0;
+}
+.bal-node.self { border-color: var(--w4f-primary); }
+.bal-node.over { border-color: var(--w4f-warning); }
+.bal-mode { width: 168px; }
+.bal-src { font-size: 11px; color: var(--w4f-muted); }
+.bal-note {
+  margin: 0 0 10px; font-size: 11.5px; line-height: 1.55;
+  color: var(--w4f-muted);
+}
+.bn-top { display: flex; align-items: center; gap: 6px; flex-wrap: wrap; margin-bottom: 7px; }
+.bn-id {
+  font-family: var(--w4f-mono, ui-monospace, monospace);
+  font-size: 12px; font-weight: 700;
+  overflow: hidden; text-overflow: ellipsis; white-space: nowrap;
+}
+.bn-stats {
+  display: flex; flex-wrap: wrap; gap: 3px 12px;
+  font-size: 11.5px; color: var(--w4f-muted); margin-top: 7px;
+}
+.bn-stats b { color: var(--w4f-text); font-variant-numeric: tabular-nums; }
+.bn-score { margin-left: auto; }
+
+.bal-plan { margin-top: 11px; }
+.bp-inflight, .bp-would, .bp-waiting, .bp-ok {
+  display: flex; align-items: center; gap: 9px; flex-wrap: wrap;
+  padding: 8px 11px; border-radius: 8px;
+  font-size: 12.5px;
+  background: var(--w4f-card-2);
+  border: 1px solid var(--w4f-line);
+}
+.bp-inflight { border-color: var(--w4f-warning); }
+.bp-text { min-width: 0; overflow-wrap: anywhere; }
+
+@media (max-width: 380px) {
+  .bn-score { margin-left: 0; }
+}
 </style>

web/src/views/SettingsView.vue

@@ -167,11 +167,12 @@ onMounted(load)
 </script>
 
 <style scoped lang="scss">
+/* Let .content (App.vue) remain the single scroll container — see the note in
+ * StatusView: a page-level scroller sized by leftover flex space can collapse
+ * to zero height on narrow screens. */
 .settings-page {
-  height: 100%;
   display: flex;
   flex-direction: column;
-  overflow-y: auto;
 }
 
 .page-top {

web/src/views/StatusView.vue

@@ -496,11 +496,22 @@ onBeforeUnmount(() => {
 </script>
 
 <style scoped lang="scss">
+/* The pages inside .content must NOT be their own scrollers.
+ *
+ * App.vue already designates `.content` as the app's single scroll container
+ * (`flex:1; overflow-y:auto`). These pages were also `height:100%` +
+ * `overflow-y:auto`, producing a nested scroller whose height is the leftover
+ * space of a flex column. On a 375px phone the status page's fixed-height
+ * header (122px) plus the single-column KPI stack (565px) exceed the available
+ * 651px, so the inner `.status-body` was handed 0px while holding 1682px of
+ * content — the page looked completely frozen, since neither `.content` (which
+ * had nothing to scroll) nor the 0-height inner box could be swiped.
+ *
+ * Letting the page grow naturally and scroll in `.content` removes the whole
+ * class of bug: there is no leftover-space arithmetic left to get wrong. */
 .status-page {
-  height: 100%;
   display: flex;
   flex-direction: column;
-  overflow: hidden;
   gap: 14px;
 }
 
@@ -527,12 +538,9 @@ onBeforeUnmount(() => {
 }
 
 .status-body {
-  flex: 1;
   display: flex;
   flex-direction: column;
   gap: 18px;
-  overflow-y: auto;
-  padding-right: 4px;
 }
 
 .section { display: flex; flex-direction: column; }
@@ -720,8 +728,13 @@ onBeforeUnmount(() => {
 }
 
 @media (max-width: 380px) {
-  /* 极窄：KPI 也降为单列，否则数字会被压断 */
-  .kpis { grid-template-columns: minmax(0, 1fr); }
+  /* Keep the KPI grid at TWO columns even on the narrowest phones.
+   * Dropping to one column made four full-width cards ~565px tall, which on its
+   * own exceeded the whole viewport and (with the page-level scroller of the
+   * day) collapsed the content area to zero height. Two columns keeps each card
+   * ~160px wide — enough for the 27px number + 11.5px caption, which is all a
+   * KPI card holds. */
+  .kpis { gap: 8px; }
   .fwd-route { font-size: 12.5px; }
   .page-title { font-size: 16px; }
 }