@bolloon/bolloon-agent 0.1.35 → 0.1.36

package/dist/web/server.js

@@ -31,15 +31,15 @@ function resolveWebRoot() {
     }
     const d = _baseDirname;
     const candidates = [
-        path.join(d), // dist/web
-        path.join(d, '..', '..', 'dist', 'web'), // src/web → dist/web
+        path.join(d, '..', '..', 'dist', 'web'), // 2026-06-15 优先: src/web → dist/web (build 产物)
+        path.join(d), // dist/web (npm 跑时)
         path.join(d, '..', 'web'), // dist/ → web/ 兄弟
     ];
     for (const c of candidates) {
         if (fsSync.existsSync(path.join(c, 'index.html')))
             return c;
     }
-    return candidates[1];
+    return candidates[0];
 }
 const webRoot = resolveWebRoot();
 console.log(`[web] webRoot = ${webRoot}`);
@@ -140,7 +140,6 @@ async function saveTheme(theme, agentId) {
 }
 // ==================== Task Queue & Workflow System ====================
 const TASK_QUEUE_PATH = path.join(SHARED_SESSION_PATH, 'task-queue.json');
-const WORKFLOW_STATE_PATH = path.join(SHARED_SESSION_PATH, 'workflow-state.json');
 async function loadTaskQueue() {
     try {
         const data = await fs.readFile(TASK_QUEUE_PATH, 'utf-8');
@@ -153,33 +152,6 @@ async function loadTaskQueue() {
 async function saveTaskQueue(tasks) {
     await fs.writeFile(TASK_QUEUE_PATH, JSON.stringify(tasks, null, 2));
 }
-async function loadWorkflowState(channelId) {
-    try {
-        const data = await fs.readFile(WORKFLOW_STATE_PATH, 'utf-8');
-        const states = JSON.parse(data);
-        return states.find(s => s.channelId === channelId) || null;
-    }
-    catch {
-        return null;
-    }
-}
-async function saveWorkflowState(state) {
-    try {
-        const data = await fs.readFile(WORKFLOW_STATE_PATH, 'utf-8');
-        const states = JSON.parse(data);
-        const index = states.findIndex(s => s.channelId === state.channelId);
-        if (index >= 0) {
-            states[index] = state;
-        }
-        else {
-            states.push(state);
-        }
-        await fs.writeFile(WORKFLOW_STATE_PATH, JSON.stringify(states, null, 2));
-    }
-    catch {
-        await fs.writeFile(WORKFLOW_STATE_PATH, JSON.stringify([state], null, 2));
-    }
-}
 let isExecutingTask = false;
 let executionTaskId = null;
 async function executeTask(task, channelId) {
@@ -221,32 +193,6 @@ async function executeTask(task, channelId) {
                     result = `📝 文档总结:\n\n${summary.summary}`;
                 }
                 break;
-            case 'workflow':
-                // 执行多步骤工作流
-                if (task.steps && task.steps.length > 0) {
-                    let loopCount = 0;
-                    for (let i = 0; i < task.steps.length; i++) {
-                        // 广播循环开始
-                        loopCount++;
-                        broadcast({ type: 'workflow_loop', loopCount, content: `开始步骤 ${i + 1}/${task.steps.length}: ${task.steps[i].name}` }, channelId);
-                        task.steps[i].status = 'running';
-                        broadcast({ type: 'task_status', taskId: task.id, status: 'running', currentStep: i, totalSteps: task.steps.length }, channelId);
-                        broadcast({ type: 'workflow_step', step: `步骤 ${i + 1}`, content: `执行中: ${task.steps[i].name}` }, channelId);
-                        // 执行步骤 - 模拟流式输出
-                        for (let j = 0; j < 3; j++) {
-                            await new Promise(resolve => setTimeout(resolve, 300));
-                            broadcast({ type: 'workflow_step', step: `步骤 ${i + 1}`, content: `执行中... (${(j + 1) * 33}%)` }, channelId);
-                        }
-                        task.steps[i].status = 'completed';
-                        task.progress = Math.round(((i + 1) / task.steps.length) * 100);
-                        broadcast({ type: 'workflow_step', step: `步骤 ${i + 1}`, content: `✅ 完成: ${task.steps[i].name}` }, channelId);
-                        broadcast({ type: 'workflow_loop', loopCount, status: 'completed', content: `步骤 ${i + 1} 完成` }, channelId);
-                        broadcast({ type: 'task_status', taskId: task.id, progress: task.progress }, channelId);
-                    }
-                    result = '✅ 工作流执行完成';
-                    broadcast({ type: 'workflow_loop', loopCount, status: 'finished', content: result }, channelId);
-                }
-                break;
             default:
                 result = '未知任务类型';
         }
@@ -632,7 +578,7 @@ async function handleV3P2PMessage(parsed, conn, comm) {
             const { getMinimax } = await import('../constraints/index.js');
             const llm = getMinimax();
             // v3 新增: 在 prompt 头部标记"这是远端访客", 让 AI 知道对方不是自己 owner
-            const visitorHint = `[系统上下文] 消息来源: 远端访客 (P2P 连接, publicKey=${senderKey.substring(0, 12)}...). 对方不是你 owner, 是通过 P2P 网络访问你这个 channel 的合作者. 称呼对方时可用 "远端访客" / "朋友" / "合作者", 不要叫 "主人".\n\n`;
+            const visitorHint = `[系统上下文] 消息来源: 远端访客 (P2P 连接, publicKey=${senderKey.substring(0, 12)}...). 对方不是你 owner, 是通过 P2P 网络访问你这个 channel 的合作者. 称呼对方时可用 "远端访客" / "朋友" / "合作者", 不要叫 "用户".\n\n`;
             // v3 新增: 也注入 channel 目录给 LLM (B 的 channel 也可以 @-mention 其他)
             let dirHint = '';
             const localChannels = (await loadChannels()).filter(c => c.id !== channelId);
@@ -654,7 +600,10 @@ async function handleV3P2PMessage(parsed, conn, comm) {
                 }
                 dirHint += '语法: 在回复中写 "@渠道名 我要说的话" 即可. 消息会持久化到目标 channel 的 session.\n\n';
             }
-            const fullPrompt = `${visitorHint}${dirHint}${judgmentHint}${text}`;
+            // 2026-06-15: P2P 远端访客路径也用显式 marker 包裹 text
+            // 2026-06-15 二次修: 把 text 放在最前 (与主路径 server.ts:1868 对齐),
+            //   避免 LLM 被 judgmentHint 末尾的 "..." 误判为整个 input 被截断
+            const fullPrompt = `【本轮用户请求】\n${text}\n【请求结束】\n\n${visitorHint}${dirHint}${judgmentHint}\n`;
             let fullResponse = '';
             // v3 新增: 流式 token 节流推给 B — 让 B 看到过程
             let lastFlushAt = 0;
@@ -1475,6 +1424,35 @@ export async function createWebServer(port = 3000, options = {}) {
                     broadcast({ type: 'workflow_step', step: event.tool || '系统', content: event.content }, channelId);
                     console.log(`[SSE 广播] workflow_step: step=${event.tool}, content="${event.content?.substring(0, 80)}..."`);
                 }
+                else if (event.type === 'step_start' || event.type === 'step_done' || event.type === 'step_error') {
+                    // 2026-06-15: 步骤状态机事件 — 原样转发 (前端 step-timeline 组件订阅)
+                    broadcast({
+                        type: event.type,
+                        tool: event.tool,
+                        content: event.content,
+                        success: event.success,
+                        output: event.output,
+                        error: event.error,
+                        args: event.args,
+                    }, channelId);
+                    // 2026-06-16: 累积 step 到 runState, 供 /api/loop/inspect 读取
+                    try {
+                        if (event.type === 'step_done' || event.type === 'step_error') {
+                            const rs = channelRunState.get(channelId);
+                            if (rs) {
+                                if (!rs.lastSteps)
+                                    rs.lastSteps = [];
+                                rs.lastSteps.push({
+                                    name: String(event.tool || event.content || 'step').slice(0, 60),
+                                    status: event.type === 'step_error' ? 'failed' : (event.success === false ? 'failed' : 'ok'),
+                                    durationMs: typeof event.durationMs === 'number' ? event.durationMs : undefined,
+                                    output: event.output ? String(event.output).slice(0, 800) : (event.error ? String(event.error).slice(0, 800) : undefined),
+                                });
+                            }
+                        }
+                    }
+                    catch { /* non-fatal */ }
+                }
                 else if (event.type === 'error') {
                     broadcast({ type: 'error', content: event.content }, channelId);
                 }
@@ -1485,7 +1463,7 @@ export async function createWebServer(port = 3000, options = {}) {
             if (realChannelDid)
                 contextHint += `[系统上下文] 当前频道名称: ${realChannelName}, 你的真实 DID: ${realChannelDid}\n`;
             // v3 新增: 标识发送方 — 让 AI 分清内部 owner vs 远端访客
-            contextHint += `[系统上下文] 消息来源: 本地 (channel 内部 owner / 此机器上的用户). 称呼对方时用 "你" 或 "主人" 即可.\n`;
+            contextHint += `[系统上下文] 消息来源: 本地 (channel 内部 owner / 此机器上的用户). 称呼对方时用 "你" 或 "用户" 即可.\n`;
             if (boundWalletAddress) {
                 contextHint += `[系统上下文] 已绑定的加密钱包地址: ${boundWalletAddress}。当用户授权或启用自动工具调用时, 可使用该地址发起链上操作。\n`;
             }
@@ -1632,7 +1610,13 @@ export async function createWebServer(port = 3000, options = {}) {
             if (contextHint)
                 contextHint += '\n';
             try {
-                fullResponse = await agent.promptStream(contextHint + text, streamCallback, runState.abortController?.signal, channelId);
+                // 2026-06-15: 把 user text 单独 marker 包起来, LLM 不会被 8K+ 的 system context 吞掉
+                //   (之前 contextHint + text 拼成一整段当 user role, 24 字符的 user input 埋在 8K+ 里看不出)
+                //   修法: contextHint 当 "背景信息", text 当 "本轮用户请求" — 显式 marker 让 LLM 区分
+                // 2026-06-15 二次修: 把 text 放在最前 (LLM 看到 input 第一眼是 user text, 不会被 judgmentHint 末尾
+                //   的 "..." 误判为整个 input 截断)
+                const markedPrompt = `【本轮用户请求】\n${text}\n【请求结束】\n\n${contextHint}`;
+                fullResponse = await agent.promptStream(markedPrompt, streamCallback, runState.abortController?.signal, channelId);
             }
             catch (err) {
                 // abort 抛错: 保留已输出的部分 (fullResponse 可能是空字符串)
@@ -1747,7 +1731,7 @@ export async function createWebServer(port = 3000, options = {}) {
     function getOrCreateRunState(channelId) {
         let s = channelRunState.get(channelId);
         if (!s) {
-            s = { running: false, queue: [], abortController: null };
+            s = { running: false, queue: [], abortController: null, lastSteps: [], lastSummary: '', lastFinalReply: '', lastTokens: {} };
             channelRunState.set(channelId, s);
         }
         return s;
@@ -2463,13 +2447,27 @@ export async function createWebServer(port = 3000, options = {}) {
                 else if (event.type === 'status' || event.type === 'tool') {
                     broadcast({ type: 'status', tool: event.tool, content: event.content }, channelId);
                 }
+                else if (event.type === 'step_start' || event.type === 'step_done' || event.type === 'step_error') {
+                    // 2026-06-15: 步骤状态机事件 — 原样转发
+                    broadcast({
+                        type: event.type,
+                        tool: event.tool,
+                        content: event.content,
+                        success: event.success,
+                        output: event.output,
+                        error: event.error,
+                        args: event.args,
+                    }, channelId);
+                }
                 else if (event.type === 'error') {
                     broadcast({ type: 'error', content: event.content }, channelId);
                 }
             };
             // 重新生成时只发送用户消息 (v3: 同时注入 channel 绑定的判断力)
+            // 2026-06-15: 同 /message 路径, 用显式 marker 包裹 userMessage, 避免 LLM 把它当背景信息
             const regenHint = await buildJudgmentHint(channel, channelId);
-            fullResponse = await agent.promptStream(regenHint + userMessage, streamCallback, undefined, channelId);
+            const markedRegen = `${regenHint}\n\n【本轮用户请求】\n${userMessage}\n【请求结束】\n`;
+            fullResponse = await agent.promptStream(markedRegen, streamCallback, undefined, channelId);
             broadcast({ type: 'ai', content: fullResponse }, channelId);
             // 更新 session
             const existingSession = await loadSession(channelId, currentSessionId);
@@ -2639,42 +2637,6 @@ export async function createWebServer(port = 3000, options = {}) {
             res.status(500).json({ error: err.message });
         }
     });
-    // 创建并执行工作流
-    app.post('/api/workflow', async (req, res) => {
-        try {
-            const { channelId, title, steps } = req.body;
-            if (!channelId || !steps || !Array.isArray(steps)) {
-                return res.status(400).json({ error: 'channelId and steps required' });
-            }
-            const tasks = await loadTaskQueue();
-            const task = {
-                id: `wf_${Date.now()}_${Math.random().toString(36).substring(2, 8)}`,
-                type: 'workflow',
-                title: title || '工作流',
-                description: `包含 ${steps.length} 个步骤的工作流`,
-                status: 'pending',
-                progress: 0,
-                createdAt: new Date().toISOString(),
-                updatedAt: new Date().toISOString(),
-                steps: steps.map((s, i) => ({
-                    id: `step_${i}`,
-                    name: s,
-                    status: 'pending'
-                })),
-                currentStep: 0
-            };
-            tasks.push(task);
-            await saveTaskQueue(tasks);
-            // 自动开始执行
-            if (!isExecutingTask) {
-                executeTask(task, channelId);
-            }
-            res.json({ ok: true, task });
-        }
-        catch (err) {
-            res.status(500).json({ error: err.message });
-        }
-    });
     // ==================== LLM 配置 API ====================
     // 获取所有 LLM 配置
     app.get('/api/llm-config', async (req, res) => {
@@ -3827,7 +3789,7 @@ export async function createWebServer(port = 3000, options = {}) {
             res.status(500).json({ error: err.message });
         }
     });
-    // 主人审阅: 批准 draft
+    // 用户审阅: 批准 draft
     app.post('/api/chat/approve', async (req, res) => {
         try {
             const { messageId, peerDID, finalText } = req.body || {};
@@ -3841,7 +3803,7 @@ export async function createWebServer(port = 3000, options = {}) {
             res.status(500).json({ error: err.message });
         }
     });
-    // 主人审阅: 丢弃 draft
+    // 用户审阅: 丢弃 draft
     app.post('/api/chat/dismiss', async (req, res) => {
         try {
             const { messageId, peerDID } = req.body || {};
@@ -4454,6 +4416,42 @@ export async function createWebServer(port = 3000, options = {}) {
             res.status(500).json({ error: err.message });
         }
     });
+    // ============================================================
+    // system-prompt health (P-Action 2 — Harness Gardening)
+    // 返回每层 lifecycle 状态: ok | stale | overdue-review | missing-frontmatter | dynamic
+    // query: ?activeOnly=1 → 只返回当前 context 激活的层
+    // ============================================================
+    app.get('/api/prompt/health', async (req, res) => {
+        try {
+            const { listLayers } = await import('../llm/system-prompt/registry.js');
+            const { evaluateLayers, markActive } = await import('../llm/system-prompt/health.js');
+            const all = listLayers();
+            const baseReport = evaluateLayers(all);
+            // 如果 query 里有 activeOnly, 跑一次 assembleSystemPrompt 拿激活列表
+            if (String(req.query.activeOnly ?? '') === '1') {
+                const { assembleSystemPrompt } = await import('../llm/system-prompt/registry.js');
+                const channel = String(req.query.channel ?? 'local');
+                const role = req.query.role;
+                const tool = req.query.tool;
+                try {
+                    const r = await assembleSystemPrompt({ channel, role, tool });
+                    const activeIds = new Set(r.layerIds);
+                    res.json(markActive(baseReport, activeIds));
+                }
+                catch (err) {
+                    console.warn('[prompt-health] assembleSystemPrompt failed (silent, returning base report):', err);
+                    res.json(baseReport);
+                }
+            }
+            else {
+                res.json(baseReport);
+            }
+        }
+        catch (err) {
+            console.error('[prompt-health] failed:', err);
+            res.status(500).json({ error: err.message });
+        }
+    });
     // 自适应接受/拒绝: 写 evolution.jsonl 留痕, 接受时同时 patch judgments.json
     // body: { action: 'accept'|'reject'|'revert', suggestion, appliedPatch? }
     // query: ?auto=1  → 类 B 自动路径, 受 auto-evolve-policy 网关保护
@@ -4919,25 +4917,192 @@ export async function createWebServer(port = 3000, options = {}) {
         const { getEventHistory } = await import('../heartbeat/self-improve-bus.js');
         res.json(getEventHistory());
     });
+    // ============================================================
+    // Permission Mode (P2.2 — UI 暴露开关)
+    // 优先级: 运行时 session 覆盖 > env BOLLOON_PERM_MODE > 'default'
+    // 运行时覆盖存在 ~/.bolloon/sessions/permission-mode.json, 每次 promptStream 入口读取
+    // ============================================================
+    const PERM_MODE_FILE = path.join(process.env.HOME || os.homedir() || '/tmp', '.bolloon', 'sessions', 'permission-mode.json');
+    function readPermModeOverride() {
+        try {
+            // 同步读, 文件很小
+            const raw = fsSync.readFileSync(PERM_MODE_FILE, 'utf-8');
+            const obj = JSON.parse(raw);
+            if (obj && typeof obj.mode === 'string') {
+                return { mode: obj.mode, ts: obj.ts || new Date().toISOString() };
+            }
+        }
+        catch { /* 不存在 = 无 override */ }
+        return null;
+    }
+    function writePermModeOverride(mode) {
+        try {
+            const dir = path.dirname(PERM_MODE_FILE);
+            fsSync.mkdirSync(dir, { recursive: true });
+            fsSync.writeFileSync(PERM_MODE_FILE, JSON.stringify({ mode, ts: new Date().toISOString() }, null, 2), 'utf-8');
+        }
+        catch (err) {
+            console.warn('[server] writePermModeOverride failed:', err);
+        }
+    }
+    // 读当前生效 mode (runtime override > env > default)
+    app.get('/api/permission-mode', async (_req, res) => {
+        const { resolvePermissionMode, ALL_PERMISSION_MODES } = await import('../agents/permission-mode.js');
+        const override = readPermModeOverride();
+        const envMode = process.env.BOLLOON_PERM_MODE || null;
+        const effective = resolvePermissionMode();
+        res.json({
+            effective,
+            override: override?.mode || null,
+            overrideTs: override?.ts || null,
+            env: envMode,
+            allowed: ALL_PERMISSION_MODES,
+            description: {
+                default: '每次工具调用询问; shell 走 shell-guard',
+                acceptEdits: 'edit_*/write_* 跳过黑名单; shell 仍走 shell-guard',
+                bypassPermissions: '非 shell 全部放行; shell 永远走 shell-guard (硬约束)',
+            },
+        });
+    });
+    // 设 runtime override (存盘, 下次 promptStream 入口读取生效)
+    app.post('/api/permission-mode', async (req, res) => {
+        const { resolvePermissionMode, ALL_PERMISSION_MODES } = await import('../agents/permission-mode.js');
+        const mode = String(req.body?.mode || '');
+        if (!ALL_PERMISSION_MODES.includes(mode)) {
+            return res.status(400).json({
+                error: `Invalid mode. Allowed: ${ALL_PERMISSION_MODES.join(', ')}`,
+                allowed: ALL_PERMISSION_MODES,
+            });
+        }
+        const oldMode = readPermModeOverride()?.mode || process.env.BOLLOON_PERM_MODE || 'default';
+        writePermModeOverride(mode);
+        // 写历史 (append-only JSONL, 跟 bolloon 其他 audit 一致)
+        try {
+            const HISTORY_FILE = path.join(process.env.HOME || os.homedir() || '/tmp', '.bolloon', 'sessions', 'permission-mode-history.jsonl');
+            fsSync.mkdirSync(path.dirname(HISTORY_FILE), { recursive: true });
+            fsSync.appendFileSync(HISTORY_FILE, JSON.stringify({
+                ts: new Date().toISOString(),
+                from: oldMode,
+                to: mode,
+                source: 'api',
+            }) + '\n', 'utf-8');
+        }
+        catch { /* 历史写失败不阻塞主流程 */ }
+        console.log(`[server] permission-mode override set to "${mode}" via API (was "${oldMode}")`);
+        res.json({
+            ok: true,
+            mode,
+            previousMode: oldMode,
+            ts: new Date().toISOString(),
+            note: '新 mode 在下一次 promptStream 入口生效 (不打断当前对话)',
+        });
+    });
+    // 取消 runtime override, 回到 env 或 default
+    app.delete('/api/permission-mode', async (_req, res) => {
+        const oldMode = readPermModeOverride()?.mode || process.env.BOLLOON_PERM_MODE || 'default';
+        try {
+            if (fsSync.existsSync(PERM_MODE_FILE))
+                fsSync.unlinkSync(PERM_MODE_FILE);
+        }
+        catch (err) {
+            console.warn('[server] delete perm-mode override failed:', err);
+        }
+        // 写历史
+        try {
+            const HISTORY_FILE = path.join(process.env.HOME || os.homedir() || '/tmp', '.bolloon', 'sessions', 'permission-mode-history.jsonl');
+            fsSync.mkdirSync(path.dirname(HISTORY_FILE), { recursive: true });
+            fsSync.appendFileSync(HISTORY_FILE, JSON.stringify({
+                ts: new Date().toISOString(),
+                from: oldMode,
+                to: 'env-or-default',
+                source: 'api',
+                action: 'delete-override',
+            }) + '\n', 'utf-8');
+        }
+        catch { /* ignore */ }
+        res.json({ ok: true, note: '已删除 runtime override, 回到 env / default' });
+    });
+    // 2026-06-16: 循环检查 — GET /api/loop/inspect?channelId=...
+    // 返回最近一轮 ReAct loop 的产出: 步骤 / 工具调用 / 压缩摘要 / 最终回复 / token 用量.
+    // 用于前端 status bar 的「✓ 检查」按钮弹 modal.
+    app.get('/api/loop/inspect', async (req, res) => {
+        try {
+            const channelId = String(req.query.channelId || '');
+            if (!channelId)
+                return res.status(400).json({ error: 'channelId required' });
+            const s = channelRunState.get(channelId);
+            if (!s)
+                return res.json({ summary: '该 channel 无活跃 loop', steps: [], finalReply: '' });
+            const steps = (s.lastSteps || []).map((st) => ({
+                name: st.name || st.tool || 'step',
+                status: st.status || 'completed',
+                durationMs: st.durationMs,
+                output: st.output || st.result || '',
+            }));
+            res.json({
+                summary: s.lastSummary || (s.running ? 'loop 仍在运行中' : 'loop 已结束'),
+                steps,
+                finalReply: s.lastFinalReply || '',
+                tokens: s.lastTokens || {},
+            });
+        }
+        catch (err) {
+            res.status(500).json({ error: err.message });
+        }
+    });
+    // 历史 (类似 self-improve history, 供前端 timeline)
+    app.get('/api/permission-mode/history', async (_req, res) => {
+        const HISTORY_FILE = path.join(process.env.HOME || os.homedir() || '/tmp', '.bolloon', 'sessions', 'permission-mode-history.jsonl');
+        try {
+            if (!fsSync.existsSync(HISTORY_FILE))
+                return res.json([]);
+            const lines = fsSync.readFileSync(HISTORY_FILE, 'utf-8').split('\n').filter(Boolean);
+            const entries = lines.map((l) => {
+                try {
+                    return JSON.parse(l);
+                }
+                catch {
+                    return null;
+                }
+            }).filter(Boolean);
+            res.json(entries.slice(-50)); // 最近 50 条
+        }
+        catch (err) {
+            res.json([]);
+        }
+    });
     // 健康检查错误数 ≥ 2 -> 触发自改信号
+    // 2026-06-16: 自迭代默认关 (用户模式), 仅 BOLLOON_DEV_MODE=1 或 selfImprove=true 启动项才装 callback
     if (healthMonitor) {
-        healthMonitor.startPeriodicCheck(60000, (status) => {
-            const errorCount = Object.values(status.checks)
-                .filter((c) => c.status === 'error').length;
-            if (errorCount >= 2) {
-                import('../heartbeat/self-improve-bus.js').then(({ reportSelfImproveEvent }) => {
-                    const failedKeys = Object.entries(status.checks)
-                        .filter(([_, c]) => c.status === 'error').map(([k]) => k).join(', ');
-                    reportSelfImproveEvent({
-                        kind: 'silent-timeout',
-                        details: `健康检查有 ${errorCount} 项失败: ${failedKeys}`
+        if (selfImproveEnabled) {
+            healthMonitor.startPeriodicCheck(60000, (status) => {
+                const errorCount = Object.values(status.checks)
+                    .filter((c) => c.status === 'error').length;
+                if (errorCount >= 2) {
+                    import('../heartbeat/self-improve-bus.js').then(({ reportSelfImproveEvent }) => {
+                        const failedKeys = Object.entries(status.checks)
+                            .filter(([_, c]) => c.status === 'error').map(([k]) => k).join(', ');
+                        reportSelfImproveEvent({
+                            kind: 'silent-timeout',
+                            details: `健康检查有 ${errorCount} 项失败: ${failedKeys}`
+                        });
                     });
-                });
-            }
-        });
+                }
+            });
+        }
+        else {
+            // 用户模式: 只跑监控不打自改信号, 心跳仍工作
+            healthMonitor.startPeriodicCheck(60000);
+            console.log('[24h] Health monitor periodic check (no self-improve)');
+        }
+    }
+    // 安装自改总线 -> SSE 桥 (开发者模式才装, 用户模式靠 /api/self-improve/trigger 手动触发)
+    if (selfImproveEnabled) {
+        void installSelfImproveHook();
+    }
+    else {
+        console.log('[self-improve] 用户模式, installSelfImproveHook 跳过 (可用 POST /api/self-improve/trigger 手动触发)');
     }
-    // 安装自改总线 -> SSE 桥
-    void installSelfImproveHook();
     // 端口冲突时自动找下一个可用端口（最多 10 次），避免 EADDRINUSE 直接崩溃
     return new Promise((resolve, reject) => {
         const maxAttempts = 10;
@@ -4957,11 +5122,28 @@ export async function createWebServer(port = 3000, options = {}) {
                 console.log('服务器已监听');
                 // 安装 chat bus -> SSE 桥 (供前端 inbox UI 实时刷新)
                 void installChatBusHook();
+                // 2026-06-16: ping 改为 data: {"type":"ping"} — 之前是 SSE 注释格式 (: ping\n\n),
+                //   浏览器 EventSource 不触发 onmessage, 客户端 60s 阈值 (现已 30s) 误判死链.
+                //   改后前端 onmessage 收到 ping 就重置 lastEventTime, 真死链才 30s 后重建.
                 setInterval(() => {
                     for (const client of sseClients) {
-                        client.res.write(': ping\n\n');
+                        try {
+                            client.res.write('data: {"type":"ping"}\n\n');
+                        }
+                        catch (err) {
+                            // socket 已断, 跳过 — client 端 onerror 会触发重连
+                            console.warn('[SSE ping] write 失败, 跳过该客户端:', err.message);
+                        }
                     }
                 }, 30000);
+                // 2026-06-16: 全局捕获 socket error 事件, 避免未处理 EPIPE/ETIMEDOUT 让进程崩
+                currentServer.on('clientError', (err, socket) => {
+                    console.warn('[server] clientError:', err.code, err.message);
+                    try {
+                        socket.end();
+                    }
+                    catch { }
+                });
                 resolve({ app, server: currentServer, port: currentPort });
             });
         };