@konglx/rotom 2.21.0

package/dist/master/scheduler-handlers.js

@@ -0,0 +1,510 @@
+import { nowBeijing, toBeijing } from "../shared/time.js";
+import { extractMentions } from "../shared/mention.js";
+/**
+ * scheduler-handlers —— 定时器到点跑硬编码逻辑的 handler 注册表。
+ *
+ * scheduled_tasks.handler_key 非空时,scheduler.runOne 不走 prompt/agent 路径,
+ * 而是查这个 registry 找对应 handler 跑。handler 接收 payload(JSON 解析后) +
+ * ctx(db / hub),自行决定做什么(查 DB、发消息、建 Issue、cancel 别的 task...)。
+ *
+ * 这是"定时器模板"机制:不同业务场景注册不同 handler,共用同一套调度基础设施
+ * (scheduler tick、at-most-once、grace window、Dashboard 可见)。
+ *
+ * 当前注册的 handler:
+ *   - ask-bridge-check:ask-bridge 超时检查,见 docs/ASK_BRIDGE_GUIDE.md
+ *
+ * 未来可加:
+ *   - issue-stale-check:Issue 长时间无进展自动告警
+ *   - agent-offline-alert:agent 离线超阈值通知真人
+ */
+import { randomUUID } from "node:crypto";
+import { resolveGroupAgentWorkingDir } from "./group-paths.js";
+import { createLogger } from "../shared/logger.js";
+import { TIMER_PERSONA_NAME } from "./util/persona.js";
+const log = createLogger("mesh-scheduler-handlers");
+const registry = {};
+export function registerSchedulerHandler(key, handler) {
+    registry[key] = handler;
+    log.info(`handler registered: ${key}`);
+}
+export function getSchedulerHandler(key) {
+    return registry[key];
+}
+// ── ask-bridge-check handler ───────────────────────────────────────────────
+/**
+ * ask-bridge 定时检查。每 20s 跑一次(interval 任务),payload: { bridgeId }。
+ *
+ * 每次 tick 检查三条路径:
+ *   1. bridge 已 resolved(answered/cancelled/timed_out)→ skip,handler 应已 disable task
+ *   2. B 有回复 → mark answered + disable task。复述按 reply.mentions 分流:
+ *      - reply 已 @ asker → A 已被 raw @ 唤醒,跳过 system 复述(避免重复提醒)
+ *      - reply 未 @ asker → postSystemToGroup 复述给 A(20s poll 兜底,raw @ 没触发)
+ *   3. expires_at < now(5min 到)→ 创建升级 Issue + mark timed_out + disable task
+ *   4. 都没命中 → return ok,等下个 20s tick
+ *
+ * ws-hub 事件式 @ 检测仍在(即时 cancel + disable task),handler 的 @ 检测是兜底。
+ * 非@回复只能靠 handler 这个 20s poll 检测(20s 内必达)。
+ */
+registerSchedulerHandler("ask-bridge-check", async (payload, ctx) => {
+    const p = payload;
+    if (!p.bridgeId)
+        return { status: "error", error: "missing bridgeId in payload" };
+    const bridge = ctx.db.getAskBridge(p.bridgeId);
+    if (!bridge)
+        return { status: "error", error: `bridge ${p.bridgeId} not found` };
+    if (bridge.status !== "pending") {
+        // 已 resolved —— task 应该已被 disable。若还在跑,disable 一下兜底。
+        cancelBridgeTask(ctx, bridge.id);
+        return { status: "skipped", error: `bridge already ${bridge.status}` };
+    }
+    const now = Date.now();
+    // 1. 查 B 是否有回复(@ 或非@都算,取最新一条)
+    const reply = ctx.db.findLatestReplyForBridge(bridge);
+    if (reply) {
+        // 有回复 → mark answered + disable task。是否发 system 复述看 reply 是否已 @ asker:
+        //   - 已 @ → A 已被 raw @ 唤醒,不复述(避免重复提醒,西花 06-28 反馈)
+        //   - 未 @ → 走 system 复述唤醒 A(20s poll 兜底路径)
+        // mentions 从 content 重抽(不信任 DB 行的 mentions 列:agent a2a_reply 入库写死 [])
+        ctx.db.markBridgeAnswered(bridge.id, reply.id);
+        cancelBridgeTask(ctx, bridge.id);
+        const replyMentions = extractMentions(reply.content);
+        if (replyMentions.includes(bridge.asker)) {
+            log.info(`bridge #${bridge.id} answered by handler: ${bridge.target} replied (msg ${reply.id}) already @ ${bridge.asker}, skip restatement`);
+            return { status: "ok" };
+        }
+        const questionContent = ctx.db.getGroupMessageContent(bridge.question_msg_id) || "(原问题已删除)";
+        const qSnippet = questionContent.slice(0, 200);
+        const replySnippet = reply.content.slice(0, 500);
+        const msg = `@${bridge.asker} ${TIMER_PERSONA_NAME} 来汇报: ${bridge.target} 在 ${reply.created_at} 回复了你之前的提问:\n` +
+            `你问: ${qSnippet}\n` +
+            `${bridge.target} 回复: ${replySnippet}\n` +
+            `\n` +
+            `**下一步**:基于这条回复继续你之前的任务。如果你不记得任务上下文(新 session),**立即跑 \`rotom group history ${bridge.group_id} --limit 10\`**,找到你问之前是谁 @ 你的、他们让你做什么,然后把 ${bridge.target} 的回复告诉那个人。`;
+        ctx.hub.postSystemToGroup(bridge.group_id, msg);
+        log.info(`bridge #${bridge.id} answered by handler: ${bridge.target} replied (msg ${reply.id}), restated to ${bridge.asker} (no @ in reply)`);
+        return { status: "ok" };
+    }
+    // 2. 超时(5min 到,无回复)→ 创建升级 Issue + disable task
+    if (bridge.expires_at < now) {
+        const issueId = randomUUID();
+        const questionContent = ctx.db.getGroupMessageContent(bridge.question_msg_id) || "(原问题已删除)";
+        const qSnippet = questionContent.slice(0, 200);
+        const createdIso = toBeijing(bridge.created_at);
+        const minutes = Math.max(1, Math.round(bridge.timeout_ms / 60_000));
+        const escalateTo = resolveEscalateTo(ctx, bridge);
+        const title = `[ask-bridge] ${bridge.target} 未回复,需升级`;
+        const description = [
+            `${TIMER_PERSONA_NAME} 等了 ${minutes} 分钟没等到 ${bridge.target},触发升级:`,
+            `你于 ${createdIso} 在群 "${bridge.group_id}" 问 ${bridge.target}:`,
+            `  "${qSnippet}"`,
+            ``,
+            escalateTo
+                ? `请去群里 @ ${escalateTo} 求救,说明:`
+                : `群里当前没有在线的真人。请在群里说明情况,等真人上线后跟进。`,
+            `- 你问的是什么`,
+            `- 等了多久`,
+            `- 你尝试过什么(如有)`,
+            ``,
+            `求救后 rotom issue complete ${issueId} 关闭此 Issue。`,
+        ].join("\n");
+        ctx.db.createIssue({
+            id: issueId,
+            groupId: bridge.group_id,
+            title,
+            description,
+            createdBy: "system:ask-bridge",
+            workingDir: resolveGroupAgentWorkingDir(ctx.db, bridge.group_id, bridge.asker),
+            assignedTo: bridge.asker,
+        });
+        ctx.db.markBridgeTimedOut(bridge.id, issueId, null);
+        cancelBridgeTask(ctx, bridge.id);
+        const pushed = ctx.hub.pushIssueAssignment(issueId, bridge.asker);
+        if (!pushed) {
+            log.warn(`bridge #${bridge.id} pushIssueAssignment to ${bridge.asker} failed (offline?)`);
+        }
+        log.info(`bridge #${bridge.id} timed_out by handler: issue ${issueId} → ${bridge.asker} (no reply, escalate)`);
+        return { status: "ok", issueId };
+    }
+    // 3. 都没命中 → 等 5min 内的下个 20s tick
+    return { status: "ok" };
+});
+const PATROL_PRIORITY_ORDER = {
+    critical: 0,
+    high: 1,
+    medium: 2,
+    low: 3,
+};
+function patrolPriorityRank(p) {
+    return PATROL_PRIORITY_ORDER[p] ?? 99;
+}
+registerSchedulerHandler("issue-patrol", async (payload, ctx) => {
+    const p = (payload ?? {});
+    const throughputCap = p.throughputCap ?? 3;
+    const candidateCap = p.candidateCap ?? 3;
+    const scanBatch = p.scanBatch ?? 10;
+    // task.group_id / task.agent_name 本应是 patrolGroupId / patrolAgentName,
+    // 但 HandlerContext 只暴露 db/hub,scheduler.runOne 没把 task 传进来。
+    // 所以 patrolGroupId/patrolAgentName 在 handler_payload 里显式携带
+    // (由 POST /api/groups 建 patrol 群时写入,PATCH /api/issues-patrol/config 可改)。
+    const { patrolGroupId, patrolAgentName } = p;
+    if (!patrolGroupId || !patrolAgentName) {
+        return { status: "error", error: "missing patrolGroupId/patrolAgentName in handler_payload" };
+    }
+    const runId = randomUUID();
+    const startedAt = nowBeijing();
+    // 1. 防 overlap:查该 patrol 群最近一次 run,若 patrol_issue 仍 in_progress 则跳过
+    const recentRuns = ctx.db.listPatrolRuns({ patrolGroupId, limit: 1 });
+    const lastRun = recentRuns[0];
+    if (lastRun && lastRun.patrol_issue_id && lastRun.status === "dispatched") {
+        const prevIssue = ctx.db.getIssueById(lastRun.patrol_issue_id);
+        if (prevIssue && prevIssue.status === "in_progress") {
+            ctx.db.createPatrolRun({
+                runId,
+                patrolGroupId,
+                startedAt,
+                inProgressCount: 0,
+                status: "skipped_overlap",
+            });
+            ctx.db.finishPatrolRun(runId, "skipped_overlap", { note: `prev patrol issue ${prevIssue.id} still in_progress` });
+            log.info(`patrol run ${runId}: skipped_overlap (prev ${prevIssue.id})`);
+            return { status: "skipped", error: "prev patrol issue in_progress" };
+        }
+    }
+    // 2. 巡检员在线检查
+    const agent = ctx.db.getAgentByName(patrolAgentName);
+    if (!agent || agent.status !== "online") {
+        ctx.db.createPatrolRun({
+            runId,
+            patrolGroupId,
+            startedAt,
+            inProgressCount: 0,
+            status: "agent_offline",
+        });
+        ctx.db.finishPatrolRun(runId, "agent_offline", { note: !agent ? "agent not found" : "agent offline" });
+        log.info(`patrol run ${runId}: agent_offline (${patrolAgentName})`);
+        return { status: "skipped", error: "patrol agent offline" };
+    }
+    // 3. 全局 in_progress 计数(排除 patrol 群自己)
+    const inProgress = ctx.db.listAllIssues("in_progress").filter((i) => i.group_id !== patrolGroupId);
+    const inProgressCount = inProgress.length;
+    if (inProgressCount >= throughputCap) {
+        ctx.db.createPatrolRun({
+            runId,
+            patrolGroupId,
+            startedAt,
+            inProgressCount,
+            status: "skipped_quota",
+        });
+        ctx.db.finishPatrolRun(runId, "skipped_quota", { note: `throughput cap reached (${inProgressCount}/${throughputCap})` });
+        log.info(`patrol run ${runId}: skipped_quota (in_progress=${inProgressCount})`);
+        return { status: "ok" };
+    }
+    // 4. 取候选 open issue(排除 patrol 群自己),按 priority 排
+    const slots = candidateCap - inProgressCount;
+    if (slots <= 0) {
+        ctx.db.createPatrolRun({
+            runId,
+            patrolGroupId,
+            startedAt,
+            inProgressCount,
+            status: "skipped_quota",
+        });
+        ctx.db.finishPatrolRun(runId, "skipped_quota", { note: `no slots (in_progress=${inProgressCount}, cap=${candidateCap})` });
+        log.info(`patrol run ${runId}: skipped_quota (no slots)`);
+        return { status: "ok" };
+    }
+    // 候选数量:既要喂饱巡检员找满 slots 个 ready,又不能扫太多浪费 token。
+    // 取 max(slots, scanBatch/2) 上限 scanBatch。
+    const take = Math.min(scanBatch, Math.max(slots, Math.ceil(scanBatch / 2)));
+    const openIssues = ctx.db.listAllIssues("open")
+        .filter((i) => i.group_id !== patrolGroupId && i.type === "task")
+        .sort((a, b) => {
+        const rankDiff = patrolPriorityRank(a.priority) - patrolPriorityRank(b.priority);
+        if (rankDiff !== 0)
+            return rankDiff;
+        return a.created_at < b.created_at ? -1 : 1;
+    })
+        .slice(0, take);
+    if (openIssues.length === 0) {
+        ctx.db.createPatrolRun({
+            runId,
+            patrolGroupId,
+            startedAt,
+            inProgressCount,
+            status: "completed",
+        });
+        ctx.db.finishPatrolRun(runId, "completed", { scanned: 0, ready: 0, note: "no open candidates" });
+        log.info(`patrol run ${runId}: no open candidates`);
+        return { status: "ok" };
+    }
+    // 5. 取规则 skill 全文
+    const skill = ctx.db.getSkillByName("issue-patrol-rules");
+    const rulesText = skill?.content ?? "(rules skill not found;fall back to default judgment)";
+    // 6. 组装 patrol issue
+    const candidateList = openIssues.map((iss, idx) => {
+        return [
+            `### 候选 ${idx + 1}: ${iss.title}`,
+            `- issue_id: ${iss.id}`,
+            `- group_id: ${iss.group_id}`,
+            `- priority: ${iss.priority}`,
+            `- slash_command: ${iss.slash_command ?? "(无)"}`,
+            `- working_dir: ${iss.working_dir ?? "(默认)"}`,
+            `- created_at: ${iss.created_at}`,
+            `- description:`,
+            iss.description?.slice(0, 800) ?? "(空)",
+        ].join("\n");
+    }).join("\n\n");
+    const title = `[巡检] ${new Date().toLocaleString("zh-CN", { hour12: false })}`;
+    const description = [
+        `# Issue 巡检任务`,
+        ``,
+        `你是巡检员。本轮全局 in_progress=${inProgressCount}(cap ${throughputCap}),剩余 slot=${slots}。`,
+        `下面是 ${openIssues.length} 个候选 open issue,请逐个判断"是否可以直接认领开工"。`,
+        ``,
+        `## 巡检规则`,
+        ``,
+        rulesText,
+        ``,
+        `## 候选列表`,
+        ``,
+        candidateList,
+        ``,
+        `## 输出要求`,
+        ``,
+        `**不要**认领、分配、或操作任何候选 issue。只输出判断。`,
+        `result 字段必须是 JSON 数组,每条:`,
+        "```json",
+        `{`,
+        `  "issue_id": "<候选 id>",`,
+        `  "verdict": "ready" | "not_ready" | "uncertain",`,
+        `  "rule_matched": "<命中的规则名>",`,
+        `  "rationale": "<一句话理由>"`,
+        `}`,
+        "```",
+        `找满 ${slots} 个 ready 就可以停(不要继续往后扫)。`,
+    ].join("\n");
+    // 7. 派 issue + 写 run
+    const issueId = randomUUID();
+    ctx.db.createIssue({
+        id: issueId,
+        groupId: patrolGroupId,
+        title,
+        description,
+        priority: "medium",
+        createdBy: "system:issue-patrol",
+        workingDir: resolveGroupAgentWorkingDir(ctx.db, patrolGroupId, patrolAgentName),
+        assignedTo: patrolAgentName,
+    });
+    ctx.db.createPatrolRun({
+        runId,
+        patrolGroupId,
+        patrolIssueId: issueId,
+        startedAt,
+        inProgressCount,
+        status: "dispatched",
+    });
+    const pushed = ctx.hub.pushIssueAssignment(issueId, patrolAgentName);
+    if (!pushed) {
+        ctx.db.finishPatrolRun(runId, "error", { note: "pushIssueAssignment failed" });
+        log.warn(`patrol run ${runId}: pushIssueAssignment to ${patrolAgentName} failed`);
+        return { status: "error", error: "pushIssueAssignment failed", issueId };
+    }
+    log.info(`patrol run ${runId}: dispatched issue ${issueId} → ${patrolAgentName} (${openIssues.length} candidates, slots=${slots})`);
+    return { status: "ok", issueId };
+});
+registerSchedulerHandler("link-patrol", async (payload, ctx) => {
+    const p = (payload ?? {});
+    const scanBatch = p.scanBatch ?? 20;
+    const { patrolGroupId, patrolAgentName } = p;
+    if (!patrolGroupId || !patrolAgentName) {
+        return { status: "error", error: "missing patrolGroupId/patrolAgentName in handler_payload" };
+    }
+    const runId = randomUUID();
+    const startedAt = nowBeijing();
+    // 1. 防 overlap:查该 patrol-link 群最近一次 run,patrol_issue 仍 in_progress 跳过
+    const recentRuns = ctx.db.listLinkPatrolRuns({ patrolGroupId, limit: 1 });
+    const lastRun = recentRuns[0];
+    if (lastRun && lastRun.patrol_issue_id && lastRun.status === "dispatched") {
+        const prevIssue = ctx.db.getIssueById(lastRun.patrol_issue_id);
+        if (prevIssue && prevIssue.status === "in_progress") {
+            ctx.db.createLinkPatrolRun({
+                runId,
+                patrolGroupId,
+                startedAt,
+                status: "skipped",
+            });
+            ctx.db.finishLinkPatrolRun(runId, "skipped", { note: `prev patrol issue ${prevIssue.id} still in_progress` });
+            log.info(`link-patrol run ${runId}: skipped_overlap (prev ${prevIssue.id})`);
+            return { status: "skipped", error: "prev link-patrol issue in_progress" };
+        }
+    }
+    // 2. 巡检员在线检查
+    const agent = ctx.db.getAgentByName(patrolAgentName);
+    if (!agent || agent.status !== "online") {
+        ctx.db.createLinkPatrolRun({
+            runId,
+            patrolGroupId,
+            startedAt,
+            status: "agent_offline",
+        });
+        ctx.db.finishLinkPatrolRun(runId, "agent_offline", { note: !agent ? "agent not found" : "agent offline" });
+        log.info(`link-patrol run ${runId}: agent_offline (${patrolAgentName})`);
+        return { status: "skipped", error: "link-patrol agent offline" };
+    }
+    // 3. 取候选未分类链接
+    const candidates = ctx.db.listUnclassifiedLinks(scanBatch);
+    if (candidates.length === 0) {
+        ctx.db.createLinkPatrolRun({
+            runId,
+            patrolGroupId,
+            startedAt,
+            status: "completed",
+        });
+        ctx.db.finishLinkPatrolRun(runId, "completed", { classified: 0, note: "no unclassified links" });
+        log.info(`link-patrol run ${runId}: no unclassified links`);
+        return { status: "ok" };
+    }
+    // 4. 拉 memory few-shot:agent_visible=1 + tags 含 link_classification,scope=global
+    const rules = ctx.db.listMemory({
+        scope: "global",
+        tags: ["link_classification"],
+        agentVisible: 1,
+        limit: 20,
+    });
+    // 5. 拉 skill 全文
+    const skill = ctx.db.getSkillByName("link-patrol-rules");
+    const rulesText = skill?.content ?? "(link-patrol-rules skill not found; fall back to default categories)";
+    // 6. 拼 description
+    const candidateList = candidates.map((l, idx) => {
+        return [
+            `### 候选 ${idx + 1}: ${l.host}`,
+            `- link_id: ${l.id}`,
+            `- url_raw: ${l.url_raw}`,
+            `- 首次上下文: ${l.first_context || "(无)"}`,
+            `- 最后见到: ${l.last_seen_at}`,
+        ].join("\n");
+    }).join("\n\n");
+    const fewShotBlock = rules.length > 0
+        ? [
+            "## 历史分类经验(few-shot,memory 自动积累)",
+            "复用这些规则;若 host 在下列出现过,优先沿用旧分类:",
+            ...rules.map((r) => `- ${r.key}: ${r.summary ?? "(无 summary)"}`),
+        ].join("\n")
+        : "(暂无历史经验,这是首轮)";
+    const title = `[链接分类] ${new Date().toLocaleString("zh-CN", { hour12: false })}`;
+    const description = [
+        "# 链接分类任务",
+        ``,
+        `本轮有 ${candidates.length} 条待分类链接。按规则给出 category + tags + title + rationale。`,
+        ``,
+        "## 分类规则",
+        "",
+        rulesText,
+        "",
+        fewShotBlock,
+        "",
+        "## 候选链接",
+        "",
+        candidateList,
+        "",
+        "## 输出要求",
+        "",
+        "issue `result` 字段必须是 JSON 数组,用 markdown ```json 代码块包裹,每条:",
+        "```json",
+        "{",
+        '  "link_id": "<uuid>",',
+        '  "category": "reference | code | tool | article | paper | discussion | issue-tracker | media | other",',
+        '  "tags": ["keyword1", "keyword2"],',
+        '  "title": "<人类可读标题>",',
+        '  "rationale": "<一句话: host + path 模式 → category 推断理由>"',
+        "}",
+        "```",
+    ].join("\n");
+    // 7. 派 issue
+    const issueId = randomUUID();
+    ctx.db.createIssue({
+        id: issueId,
+        groupId: patrolGroupId,
+        title,
+        description,
+        priority: "medium",
+        createdBy: "system:link-patrol",
+        workingDir: resolveGroupAgentWorkingDir(ctx.db, patrolGroupId, patrolAgentName),
+        assignedTo: patrolAgentName,
+    });
+    ctx.db.createLinkPatrolRun({
+        runId,
+        patrolGroupId,
+        patrolIssueId: issueId,
+        startedAt,
+        candidatesScanned: candidates.length,
+        status: "dispatched",
+    });
+    const pushed = ctx.hub.pushIssueAssignment(issueId, patrolAgentName);
+    if (!pushed) {
+        ctx.db.finishLinkPatrolRun(runId, "error", { note: "pushIssueAssignment failed" });
+        log.warn(`link-patrol run ${runId}: pushIssueAssignment to ${patrolAgentName} failed`);
+        return { status: "error", error: "pushIssueAssignment failed", issueId };
+    }
+    log.info(`link-patrol run ${runId}: dispatched issue ${issueId} → ${patrolAgentName} (${candidates.length} candidates)`);
+    return { status: "ok", issueId };
+});
+/** cancel 对应 bridgeId 的 ask-bridge scheduled_task(disable,保留在列表里做审计)。
+ *  按 handler_key + handler_payload 查,不依赖 name(name 已改成"星期五 · …"友好文案)。 */
+function cancelBridgeTask(ctx, bridgeId) {
+    const task = ctx.db.findAskBridgeScheduledTask(bridgeId);
+    if (task && task.enabled) {
+        ctx.db.disableScheduledTask(task.id);
+        log.info(`bridge ${bridgeId}: cancelled schedule task #${task.id}`);
+    }
+}
+/**
+ * 解析超时升级的求救对象。
+ *
+ * 优先级:
+ *   1. bridge.escalate_to 显式指定 —— 直接用
+ *   2. 群里在线真人(群成员 profile override 或全局 agent profile 里 category=真人,
+ *      且 status=online,且不是 asker 自己)—— 返回 "名字1 或 名字2"
+ *   3. 没有在线真人 —— 返回 null(调用方给另一套文案,不让 asker LLM 自己挑,
+ *      否则它会乱猜 agent 当真人,见 issue bcb92df5)
+ */
+function resolveEscalateTo(ctx, bridge) {
+    if (bridge.escalate_to)
+        return bridge.escalate_to;
+    const members = ctx.db.getGroupMembers(bridge.group_id);
+    const agentByName = new Map();
+    for (const a of ctx.db.listAgents()) {
+        agentByName.set(a.name, a);
+    }
+    const humans = [];
+    for (const m of members) {
+        if (m.agent_name === bridge.asker)
+            continue;
+        const agent = agentByName.get(m.agent_name);
+        if (!agent || agent.status !== "online")
+            continue;
+        // category 优先取群成员 profile override,回落到全局 agent profile
+        let category;
+        if (m.profile) {
+            try {
+                const p = JSON.parse(m.profile);
+                if (typeof p.category === "string")
+                    category = p.category;
+            }
+            catch { /* fall through */ }
+        }
+        if (!category && agent.profile) {
+            try {
+                const p = JSON.parse(agent.profile);
+                if (typeof p.category === "string")
+                    category = p.category;
+            }
+            catch { /* ignore */ }
+        }
+        if (category === "真人")
+            humans.push(m.agent_name);
+    }
+    if (humans.length === 0)
+        return null;
+    return humans.join(" 或 ");
+}