@sleep2agi/commhub-server 0.9.0-preview.86 → 0.9.0-preview.87

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sleep2agi/commhub-server",
-  "version": "0.9.0-preview.86",
+  "version": "0.9.0-preview.87",
   "description": "CommHub Server — AI Agent communication hub with MCP protocol, multi-network isolation, user auth, and MCP tools (17 collaboration-core + node/provider ops tools; authoritative list: docs-site/docs/api/mcp-tools.md).",
   "type": "module",
   "main": "src/index.ts",

package/src/model-auth-notify-http.test.ts

@@ -0,0 +1,181 @@
+// #462 —— 节点模型登录失效 → 通知节点主人一次。
+//
+// 钉住:
+//   1. 翻转判定(纯函数):进入 revoked/expired 发一次;坏→坏不再发(含 revoked⇄expired);回到 ok 才重新上膛;
+//      unknown / 没报既不发也不上膛。
+//   2. 真 report_status(节点令牌):ok→expired→expired→ok→expired = 恰好 2 条,只给节点主人;
+//      网络 owner 和别的成员 0 条。正文带别名、原因、`codex login` 修法,不教人拷别人的 auth.json。
+//   3. 别的节点 / 用户令牌替它报坏状态 → 不发(与 #448 同一条「只有本节点令牌能替自己说健康」)。
+//   4. Hub 重启(内存标记清空)后第一眼看到坏状态:主人还有没读的同类消息 → 不再发;读过了 → 发。
+//
+// 跑法:cd server && COMMHUB_DB=/tmp/x.db bun test src/model-auth-notify-http.test.ts
+//       PG:COMMHUB_TEST_PG_URL=… COMMHUB_PG_EXPERIMENTAL=1(tests/test2123-hub-postgres-ladder 里注册)
+import { afterAll, beforeAll, beforeEach, describe, expect, test } from "bun:test";
+import { mkdtempSync } from "node:fs";
+import { join } from "node:path";
+import { tmpdir } from "node:os";
+import { Client } from "@modelcontextprotocol/sdk/client/index.js";
+import { InMemoryTransport } from "@modelcontextprotocol/sdk/inMemory.js";
+import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
+import { addNetworkMember, createNetworkTokenForNode, register } from "./auth.js";
+import { db } from "./db.js";
+import { registerTools } from "./tools.js";
+import { clearNodeHealthStore } from "./node-health-store.js";
+import { __resetModelAuthNoticesForTest, decideModelAuthNotice, MODEL_AUTH_NOTICE_KIND, modelAuthNoticeText } from "./model-auth-notify.js";
+
+const dir = mkdtempSync(join(tmpdir(), "anet-model-auth-notify-"));
+const activeDbPath = process.env.COMMHUB_DB ?? (process.env.COMMHUB_TEST_PG_URL ? "postgres" : undefined);
+if (!activeDbPath) throw new Error("model-auth-notify requires COMMHUB_DB (or COMMHUB_TEST_PG_URL) before module import");
+
+const PW = "ModelAuthNotify123!";
+const stamp = Date.now();
+const NODE = `ma-node-${stamp}`;
+const NODE_ID = `n_ma_node_${stamp}`;
+const PEER = `ma-peer-${stamp}`;
+const PEER_ID = `n_ma_peer_${stamp}`;
+let netId = "";
+let bossId = "", noraId = "", ottoId = "";
+let nodeTokenId = "", peerTokenId = "";
+
+async function connectAs(alias: string, userId: string, tokenId: string, isNetworkToken = true) {
+  const s = new McpServer({ name: "model-auth-notify", version: "1" });
+  registerTools(s, undefined, netId, userId, alias, isNetworkToken, tokenId);
+  const client = new Client({ name: "model-auth-notify-client", version: "1" });
+  const [ct, st] = InMemoryTransport.createLinkedPair();
+  await s.connect(st);
+  await client.connect(ct);
+  const report = async (model_auth: string | undefined, asAlias = NODE, nodeId = NODE_ID) => {
+    const r: any = await client.callTool({
+      name: "report_status",
+      arguments: { resume_id: `sdk-${nodeId}`, alias: asAlias, status: "idle", node_id: nodeId, network_id: netId, health: { bridge: "ok", ...(model_auth ? { model_auth } : {}) } },
+    });
+    expect(r.isError === true).toBe(false);
+  };
+  return { report, close: async () => { await client.close(); await s.close(); } };
+}
+
+const notices = (userId: string) => db.all<{ message_id: string; from_session: string; title: string; content: string; severity: string; meta_json: string; acked: number }>(
+  "SELECT message_id, from_session, title, content, severity, meta_json, acked FROM user_inbox WHERE user_id = ?1 AND network_id = ?2 AND kind = ?3 ORDER BY created_at, message_id",
+  userId, netId, MODEL_AUTH_NOTICE_KIND,
+);
+const clearNotices = () => db.run("DELETE FROM user_inbox WHERE network_id = ?1 AND kind = ?2", [netId, MODEL_AUTH_NOTICE_KIND]);
+
+beforeAll(() => {
+  process.env.COMMHUB_UPLOADS_DIR = join(dir, "uploads");
+  const boss = register(`ma_boss_${stamp}`, PW);
+  expect(boss.ok).toBe(true);
+  bossId = boss.user!.user_id; netId = boss.network_id!;
+  const nora = register(`ma_nora_${stamp}`, PW);
+  const otto = register(`ma_otto_${stamp}`, PW);
+  noraId = nora.user!.user_id; ottoId = otto.user!.user_id;
+  addNetworkMember(netId, noraId, "member", bossId, { agentAccess: "all" });
+  addNetworkMember(netId, ottoId, "member", bossId, { agentAccess: "all" });
+  // 节点主人是 nora(她铸的节点令牌 → nodes.owner_user_id),不是网络 owner boss。
+  const minted = createNetworkTokenForNode(noraId, netId, NODE, NODE_ID);
+  expect(minted.ok).toBe(true);
+  nodeTokenId = minted.token_id!;
+  const peer = createNetworkTokenForNode(ottoId, netId, PEER, PEER_ID);
+  expect(peer.ok).toBe(true);
+  peerTokenId = peer.token_id!;
+});
+beforeEach(() => { __resetModelAuthNoticesForTest(); clearNodeHealthStore(); clearNotices(); });
+afterAll(() => { clearNotices(); __resetModelAuthNoticesForTest(); clearNodeHealthStore(); });
+
+describe("#462 transition logic", () => {
+  test("bad once per entry, re-armed only by ok", () => {
+    const seq: Array<Parameters<typeof decideModelAuthNotice>[1]> = ["ok", "expired", "expired", "revoked", "unknown", undefined, "expired", "ok", "unknown", "revoked", "revoked"];
+    let mark: ReturnType<typeof decideModelAuthNotice>["mark"];
+    const fired: boolean[] = [];
+    for (const s of seq) { const d = decideModelAuthNotice(mark, s); fired.push(d.notify); mark = d.mark; }
+    expect(fired).toEqual([false, true, false, false, false, false, false, false, false, true, false]);
+  });
+
+  test("first sight is bad → notify; unknown/missing never arms", () => {
+    expect(decideModelAuthNotice(undefined, "revoked")).toEqual({ notify: true, mark: "notified" });
+    expect(decideModelAuthNotice(undefined, "unknown")).toEqual({ notify: false, mark: undefined });
+    expect(decideModelAuthNotice("notified", "unknown")).toEqual({ notify: false, mark: "notified" });
+    expect(decideModelAuthNotice("notified", undefined)).toEqual({ notify: false, mark: "notified" });
+    expect(decideModelAuthNotice("notified", "ok")).toEqual({ notify: false, mark: "armed" });
+  });
+
+  test("text names the node, the cause and the per-node re-login; never says to copy credentials", () => {
+    for (const s of ["revoked", "expired"] as const) {
+      const { title, text } = modelAuthNoticeText("node-a", s);
+      expect(title).toBe("节点登录失效");
+      expect(text).toContain("节点 node-a");
+      expect(text).toContain("codex login");
+      expect(text).toContain("CODEX_HOME=");
+      expect(text).toContain("不要拷别的节点的 auth.json");
+    }
+    expect(modelAuthNoticeText("node-a", "revoked").text).toContain("作废");
+    expect(modelAuthNoticeText("node-a", "expired").text).toContain("过期");
+  });
+});
+
+describe("#462 report_status → owner notice", () => {
+  test("ok → expired → expired → ok → expired = exactly 2 notices, owner only", async () => {
+    const n = await connectAs(NODE, noraId, nodeTokenId);
+    try {
+      for (const s of ["ok", "expired", "expired", "ok", "expired"]) await n.report(s);
+    } finally { await n.close(); }
+    const mine = notices(noraId);
+    expect(mine.length).toBe(2);
+    for (const m of mine) {
+      expect(m.from_session).toBe(NODE);
+      expect(m.title).toBe("节点登录失效");
+      expect(m.severity).toBe("warning");
+      expect(m.content).toContain(`节点 ${NODE}`);
+      expect(m.content).toContain("codex login");
+      expect(JSON.parse(m.meta_json)).toEqual({ model_auth_notice: { alias: NODE, state: "expired" } });
+    }
+    expect(notices(bossId).length).toBe(0);
+    expect(notices(ottoId).length).toBe(0);
+  });
+
+  test("a node that only ever reports ok / unknown / nothing gets no notice", async () => {
+    const n = await connectAs(NODE, noraId, nodeTokenId);
+    try {
+      for (const s of ["unknown", "ok", undefined, "unknown", "ok"]) await n.report(s);
+    } finally { await n.close(); }
+    expect(notices(noraId).length).toBe(0);
+  });
+
+  test("another node or a user login cannot trigger a notice for this node", async () => {
+    const p = await connectAs(PEER, ottoId, peerTokenId);
+    try { await p.report("revoked", NODE, NODE_ID); } finally { await p.close(); }
+    const u = await connectAs("", bossId, "", false);
+    try { await u.report("revoked", NODE, NODE_ID); } finally { await u.close(); }
+    expect(notices(noraId).length).toBe(0);
+    expect(notices(bossId).length).toBe(0);
+    expect(notices(ottoId).length).toBe(0);
+  });
+
+  test("hub restart: an unread notice suppresses the repeat; once read, the next sight notifies", async () => {
+    const n = await connectAs(NODE, noraId, nodeTokenId);
+    try {
+      await n.report("revoked");
+      expect(notices(noraId).length).toBe(1);
+      __resetModelAuthNoticesForTest(); // = Hub 重启,内存标记没了
+      await n.report("revoked");
+      expect(notices(noraId).length).toBe(1);
+      db.run("UPDATE user_inbox SET acked = 1 WHERE user_id = ?1 AND network_id = ?2 AND kind = ?3", [noraId, netId, MODEL_AUTH_NOTICE_KIND]);
+      __resetModelAuthNoticesForTest();
+      await n.report("revoked");
+      expect(notices(noraId).length).toBe(2);
+      await n.report("revoked");
+      expect(notices(noraId).length).toBe(2);
+    } finally { await n.close(); }
+  });
+
+  test("owner no longer in the network → nobody is notified", async () => {
+    db.run("DELETE FROM network_members WHERE network_id = ?1 AND user_id = ?2", [netId, noraId]);
+    try {
+      const n = await connectAs(NODE, noraId, nodeTokenId);
+      try { await n.report("expired"); } finally { await n.close(); }
+      expect(notices(noraId).length).toBe(0);
+      expect(notices(bossId).length).toBe(0);
+    } finally {
+      addNetworkMember(netId, noraId, "member", bossId, { agentAccess: "all" });
+    }
+  });
+});

package/src/model-auth-notify.ts

@@ -0,0 +1,132 @@
+// #462(#448 的子项)—— 节点模型登录失效时通知节点主人。
+//
+// 现场:codex 节点的 refresh token 被作废 / 过期后,节点自己知道(report_status.health.model_auth),
+// Hub 也会拒绝给它派活(#460 node-health-guard.ts),但**没有人被告诉** —— 主人要等到某次派活被 409
+// 才发现。这里在 report_status 收下健康报告的那一刻看 model_auth 的翻转:进入 revoked / expired 时
+// 给节点主人发**一条**消息,说是哪个节点、出了什么事、怎么重新登录。
+//
+// 🔴 只检测 + 通知。不自动换号、不拷别的节点的 auth.json、不重新 stage 凭据(owner 定的边界)。
+//
+// 去重:每次「进入坏状态」只发一条;回到 ok 才重新上膛。unknown / 没报 model_auth 既不发也不上膛
+//   (一次模型调用都还没跑的节点报 unknown,不能把它当「恢复了」,否则 expired→unknown→expired 会刷屏)。
+//   revoked ⇄ expired 互相切换不算新的翻转 —— 修法是同一条命令。
+// 状态在内存里(同 node-health-store.ts);Hub 重启后第一次看到坏状态时,先查收件人有没有**还没读**的同类
+//   消息,有就不再发 —— 否则每次升级 Hub 都会给还没修好的节点主人再发一遍。
+//
+// 收件人 = 节点主人:上报令牌绑定的 nodes.owner_user_id(RFC-036,建节点时定死),没有绑定(老令牌)就取
+//   铸这个节点令牌的用户(api_tokens.user_id —— 新令牌里两者相同)。只发给这一个人;不在本网络了就不发。
+// 怎么送:user_inbox(与 send_desktop_message 同一张表、同一条 /events/users/me 推送),from_session = 节点别名,
+//   所以 App 把它显示在**与该节点的会话**里(主动消息)+ 未读角标 + 通知。不新增渠道。
+
+import { db, uuidv4 } from "./db.js";
+import { getUserNetworkRole } from "./auth.js";
+import { pushUserEvent } from "./push.js";
+import { degradedLayers } from "./node-health-guard.js";
+import type { NodeHealth } from "./node-health-store.js";
+
+export const MODEL_AUTH_NOTICE_KIND = "node_model_auth";
+const MAX_TRACKED = 5_000;
+
+type ModelAuth = NodeHealth["model_auth"];
+/** armed = 下一次进入坏状态要发;notified = 这一段坏状态已经发过。 */
+export type NoticeMark = "armed" | "notified";
+
+export const isBadModelAuth = (s: ModelAuth): s is "revoked" | "expired" => s === "revoked" || s === "expired";
+
+/**
+ * 纯函数:上一份标记 + 这一份 model_auth → 发不发、新标记。
+ * prev = undefined 表示这个 Hub 进程还没见过这个节点(调用方另查库里有没有未读的旧通知)。
+ */
+export function decideModelAuthNotice(prev: NoticeMark | undefined, state: ModelAuth): { notify: boolean; mark: NoticeMark | undefined } {
+  if (state === "ok") return { notify: false, mark: "armed" };
+  if (isBadModelAuth(state)) {
+    if (prev === "notified") return { notify: false, mark: "notified" };
+    return { notify: true, mark: "notified" };
+  }
+  return { notify: false, mark: prev }; // unknown / 没报:保持原样
+}
+
+export function modelAuthNoticeText(alias: string, state: "revoked" | "expired"): { title: string; text: string } {
+  const what = state === "revoked"
+    ? "模型登录凭据已被作废(refresh token revoked),需要重新登录"
+    : "模型登录已过期,且没能自动刷新";
+  // 修法与派发被拒时(#460)给的同一句,别出现两套说法。
+  const hint = degradedLayers({ model_auth: state })[0]?.hint ?? "";
+  return {
+    title: "节点登录失效",
+    text: `节点 ${alias} 的${what}。在它恢复之前,派给它的任务会被拒绝。\n\n怎么修:${hint}。\n登录好之后节点会自动报告恢复,不用再做别的。`,
+  };
+}
+
+const marks = new Map<string, NoticeMark>();
+const keyOf = (networkId: string, alias: string) => `${networkId}\0${alias}`;
+
+export function __resetModelAuthNoticesForTest(): void { marks.clear(); }
+
+/** 上报令牌对应的节点主人。 */
+function ownerOfNodeToken(tokenId: string, networkId: string): string | null {
+  const row = db.get<{ user_id: string | null; owner_user_id: string | null }>(
+    `SELECT t.user_id, n.owner_user_id FROM api_tokens t LEFT JOIN nodes n ON n.node_id = t.bound_node_id
+      WHERE t.token_id = ?1 AND t.network_id = ?2`,
+    tokenId, networkId,
+  );
+  return row?.owner_user_id || row?.user_id || null;
+}
+
+function hasUnreadNotice(userId: string, networkId: string, alias: string): boolean {
+  return !!db.get(
+    "SELECT 1 AS hit FROM user_inbox WHERE user_id = ?1 AND network_id = ?2 AND from_session = ?3 AND kind = ?4 AND acked = 0 LIMIT 1",
+    userId, networkId, alias, MODEL_AUTH_NOTICE_KIND,
+  );
+}
+
+/**
+ * report_status 收下一份(已验过是本节点令牌报的)健康报告之后调用。返回这次通知到的 user_id(测试用;没发 = [])。
+ * 任何异常都吞掉:这是附带的通知,不能让一份正常的 report_status 失败。
+ */
+export function noteModelAuthHealth(input: { networkId: string; alias: string; tokenId: string; health: NodeHealth }): string[] {
+  try {
+    const { networkId, alias, tokenId, health } = input;
+    const state = health.model_auth;
+    const key = keyOf(networkId, alias);
+    const prev = marks.get(key);
+    const { notify, mark } = decideModelAuthNotice(prev, state);
+    if (mark) {
+      marks.delete(key);
+      marks.set(key, mark);
+      if (marks.size > MAX_TRACKED) marks.delete(marks.keys().next().value!);
+    }
+    if (!notify || !isBadModelAuth(state)) return [];
+
+    const owner = ownerOfNodeToken(tokenId, networkId);
+    if (!owner || !getUserNetworkRole(owner, networkId)) return [];
+    // Hub 重启后的第一眼:主人还有一条没读的同类消息 → 这是同一段坏状态,不再发。
+    if (prev === undefined && hasUnreadNotice(owner, networkId, alias)) return [];
+
+    const { title, text } = modelAuthNoticeText(alias, state);
+    const messageId = `dm_auth_${uuidv4()}`;
+    const meta = { model_auth_notice: { alias, state } };
+    db.run(
+      `INSERT INTO user_inbox (message_id, network_id, user_id, from_session, kind, title, content, severity, meta_json)
+       VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7, 'warning', ?8)
+       ON CONFLICT(message_id) DO NOTHING`,
+      [messageId, networkId, owner, alias, MODEL_AUTH_NOTICE_KIND, title, text, JSON.stringify(meta)],
+    );
+    // 推送在写库之后:收到推送的 App 立刻回读得到这一行。
+    pushUserEvent(networkId, owner, {
+      type: "desktop_message",
+      message_id: messageId,
+      kind: MODEL_AUTH_NOTICE_KIND,
+      from: alias,
+      title,
+      message: text,
+      severity: "warning",
+      created_at: new Date().toISOString(),
+      meta,
+    });
+    return [owner];
+  } catch (e: any) {
+    console.error(`[model-auth-notify] ${input.alias}: ${e?.message ?? e}`);
+    return [];
+  }
+}

package/src/tools.ts

@@ -3,6 +3,7 @@ import { parseDbTimestampMs } from "./db-timestamp.js";
 import { z } from "zod/v4";
 import { nodeHealthSchema, recordNodeHealth } from "./node-health-store.js";
 import { assertNodeHealthy } from "./node-health-guard.js";
+import { noteModelAuthHealth } from "./model-auth-notify.js";
 import { parseAliasFilter } from "./alias-filter.js";
 import { createHash } from "node:crypto";
 import { db, uuidv4, logTaskEvent, chainReplyToParent, hashToken, generateId, generateNetworkToken, syncScheduledRunForTask } from "./db.js";
@@ -1062,7 +1063,11 @@ export function registerTools(server: McpServer, clientIP?: string, enforceNetwo
       const acceptedHealth = nodeHealth && callerTokenIsNetwork && callerAlias && callerAlias === effectiveAlias
         ? nodeHealth
         : undefined;
-      if (acceptedHealth) recordNodeHealth(sessionNetId, effectiveAlias, acceptedHealth);
+      if (acceptedHealth) {
+        recordNodeHealth(sessionNetId, effectiveAlias, acceptedHealth);
+        // #462 — model login went bad → tell the node's owner once (detect + notify only).
+        if (callerTokenId) noteModelAuthHealth({ networkId: sessionNetId, alias: effectiveAlias, tokenId: callerTokenId, health: acceptedHealth });
+      }
       pushEvent(effectiveAlias, {
         type: "status_update",
         alias: effectiveAlias,