@sleep2agi/commhub-server 0.9.0-preview.42 → 0.9.0-preview.44

package/src/create-node-blocked-gate.test.ts

@@ -0,0 +1,356 @@
+import { afterAll, beforeEach, describe, expect, test } from "bun:test";
+import { db } from "./db.js";
+import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
+import { registerTools } from "./tools.js";
+
+// #1353 Fix ① — hub gates create_node dispatch on the daemon's
+// self-reported can_create_nodes bit. Before this, daemon reported
+// can_create_nodes=false (via #1371 + #1377) → hub stored it → nothing
+// read it → dispatch returned {ok:true, request_id} and pushed a
+// doorbell to a daemon that had already told hub it couldn't create
+// nodes. That's the exact "在线但静默失败" case the issue title reports.
+//
+// Vincent 2026-08-28 comment on #1353:
+//   "daemon 侧确知不可用时,hub 应当**拒绝派发**并返回可操作的错误,
+//    而不是发一个注定失败的 doorbell"
+//
+// 🔴 Pre-#1371 daemons (preview.55 and earlier) don't report the field
+// at all. That must NOT fail-closed — the guard is "known-blocked",
+// not "unknown-treated-as-blocked". Tested explicitly below.
+
+const NET = "net_1353_gate";
+const USER_ID = "u_1353_admin";
+const DAEMON_ID = "node_1353_daemon";
+const DAEMON_ALIAS = "gate-daemon";
+const DAEMON_TOK = "tok_1353_daemon";
+
+interface ToolHandler { (args: any, extra?: any): Promise<{ content: Array<{ type: "text"; text: string }> }>; }
+interface Reply { ok?: boolean; error?: string; blocked_reason?: string; request_id?: string; [k: string]: unknown; }
+
+function cleanup() {
+  try { db.run("DELETE FROM node_create_requests WHERE network_id = ?1", [NET]); } catch {}
+  try { db.run("DELETE FROM audit_log WHERE network_id = ?1", [NET]); } catch {}
+  try { db.run("DELETE FROM nodes WHERE network_id = ?1", [NET]); } catch {}
+  try { db.run("DELETE FROM sessions WHERE network_id = ?1", [NET]); } catch {}
+  try { db.run("DELETE FROM api_tokens WHERE network_id = ?1", [NET]); } catch {}
+  try { db.run("DELETE FROM network_members WHERE network_id = ?1", [NET]); } catch {}
+  try { db.run("DELETE FROM networks WHERE network_id = ?1", [NET]); } catch {}
+  try { db.run("DELETE FROM users WHERE user_id = ?1", [USER_ID]); } catch {}
+}
+beforeEach(cleanup);
+afterAll(cleanup);
+
+/** #1545 —— 可选:写一条 sessions 心跳行。
+ *  默认**不写**,这样既有那 9 条测试的行为逐字不变;
+ *  只有要断言绝对年龄的用例才显式给 `heartbeatAgoMs`。 */
+function seedHeartbeat(agoMs: number) {
+  const iso = new Date(Date.now() - agoMs).toISOString();
+  db.run(
+    `INSERT INTO sessions (alias, node_id, network_id, status, last_seen_at, updated_at)
+     VALUES (?1, ?2, ?3, 'idle', ?4, ?4)`,
+    [DAEMON_ALIAS, DAEMON_ID, NET, iso],
+  );
+}
+
+function seed(snapshot: object | string | null) {
+  db.run(
+    `INSERT INTO users (user_id, username, password_hash, role, created_at)
+     VALUES (?1, ?2, 'x', 'user', datetime('now'))`,
+    [USER_ID, USER_ID],
+  );
+  db.run(
+    `INSERT OR REPLACE INTO networks (network_id, network_name, owner_id, created_at)
+     VALUES (?1, ?1, ?2, datetime('now'))`,
+    [NET, USER_ID],
+  );
+  db.run(
+    `INSERT INTO network_members (user_id, network_id, role, joined_at)
+     VALUES (?1, ?2, 'admin', datetime('now'))`,
+    [USER_ID, NET],
+  );
+  const snap = snapshot === null
+    ? null
+    : (typeof snapshot === "string" ? snapshot : JSON.stringify(snapshot));
+  db.run(
+    `INSERT INTO nodes (node_id, node_name, alias, network_id, config_snapshot, hostname, created_at, updated_at, lifecycle_state)
+     VALUES (?1, ?2, ?3, ?4, ?5, ?6, datetime('now'), datetime('now'), 'active')`,
+    [DAEMON_ID, DAEMON_ALIAS, DAEMON_ALIAS, NET, snap, `host-${DAEMON_ALIAS}`],
+  );
+  db.run(
+    `INSERT INTO api_tokens (token_id, user_id, network_id, scope, name, token_hash, expires_at, revoked_at)
+     VALUES (?1, ?2, ?3, 'network', ?4, ?5, NULL, NULL)`,
+    [DAEMON_TOK, USER_ID, NET, `node:${DAEMON_ALIAS}`, `hash_${DAEMON_TOK}`],
+  );
+}
+
+function tools(): Record<string, ToolHandler> {
+  const server = new McpServer({ name: "test", version: "0" }) as any;
+  const out: Record<string, ToolHandler> = {};
+  const origTool = server.tool.bind(server);
+  server.tool = (name: string, _desc: string, _schema: any, handler: ToolHandler) => {
+    out[name] = handler;
+    return origTool(name, _desc, _schema, handler);
+  };
+  const origRegisterTool = server.registerTool?.bind(server);
+  if (origRegisterTool) {
+    server.registerTool = (name: string, _cfg: any, handler: ToolHandler) => {
+      out[name] = handler;
+      return origRegisterTool(name, _cfg, handler);
+    };
+  }
+  registerTools(
+    server, undefined,
+    /* enforceNetworkId */ NET,
+    /* enforceUserId */ USER_ID,
+    /* callerAlias */ null,
+    /* callerTokenIsNetwork */ true,
+    /* callerTokenId */ null,
+  );
+  return out;
+}
+
+async function call(handler: ToolHandler, args: any): Promise<Reply> {
+  const r = await handler(args);
+  return JSON.parse(r.content[0].text) as Reply;
+}
+
+const spec = {
+  name: "child-1353",
+  runtime: "claude-agent-sdk",
+  model: "claude-sonnet-4-5",
+  flags: {},
+};
+
+describe("#1353 Fix ① — hub gates create_node on daemon-reported can_create_nodes", () => {
+  test("🔴 witnessed-red: blocked daemon (can_create_nodes=false + reason) → refuse with reason, NO dispatch", async () => {
+    seed({
+      role: "host_supervisor",
+      daemon_capabilities: {
+        runtimes_supported: ["claude-agent-sdk"],
+        can_create_nodes: false,
+        create_nodes_blocked_reason: "anet_bin_permission",
+      },
+    });
+    const t = tools();
+    const r = await call(t.create_node, { daemon_node_id: DAEMON_ID, node_spec: spec });
+    expect(r.ok).toBe(false);
+    expect(r.error).toBe("daemon_cannot_create_nodes");
+    expect(r.blocked_reason).toBe("anet_bin_permission");
+    // No request row inserted — hub must refuse BEFORE dispatch.
+    const rows = db.get<{ n: number }>(
+      "SELECT COUNT(*) AS n FROM node_create_requests WHERE daemon_node_id = ?1",
+      DAEMON_ID,
+    );
+    expect(rows?.n).toBe(0);
+  });
+
+  test("blocked daemon with each of the 4 reason codes surfaces that specific code", async () => {
+    const codes = ["anet_bin_identity", "anet_bin_source", "anet_bin_shape", "anet_bin_permission"];
+    for (const code of codes) {
+      cleanup();
+      seed({
+        role: "host_supervisor",
+        daemon_capabilities: {
+          runtimes_supported: ["claude-agent-sdk"],
+          can_create_nodes: false,
+          create_nodes_blocked_reason: code,
+        },
+      });
+      const r = await call(tools().create_node, { daemon_node_id: DAEMON_ID, node_spec: spec });
+      expect(r.error).toBe("daemon_cannot_create_nodes");
+      expect(r.blocked_reason).toBe(code);
+    }
+  });
+
+  test("blocked with .40-shape 'anet_bin_pin_unresolved' reason surfaces that literal (backcompat)", async () => {
+    // The Zod enum in tools.ts still accepts this coarse pre-#1377 code
+    // because agent-node@2.5.0-preview.40 daemons in the wild still emit
+    // it. If we ever swallow it into "unknown", those daemons' user
+    // messages get worse than they already are.
+    seed({
+      role: "host_supervisor",
+      daemon_capabilities: {
+        runtimes_supported: ["claude-agent-sdk"],
+        can_create_nodes: false,
+        create_nodes_blocked_reason: "anet_bin_pin_unresolved",
+      },
+    });
+    const r = await call(tools().create_node, { daemon_node_id: DAEMON_ID, node_spec: spec });
+    expect(r.blocked_reason).toBe("anet_bin_pin_unresolved");
+  });
+
+  test("blocked with NO reason field → fallback 'anet_bin_unknown' (defensive)", async () => {
+    seed({
+      role: "host_supervisor",
+      daemon_capabilities: {
+        runtimes_supported: ["claude-agent-sdk"],
+        can_create_nodes: false,
+        // no create_nodes_blocked_reason
+      },
+    });
+    const r = await call(tools().create_node, { daemon_node_id: DAEMON_ID, node_spec: spec });
+    expect(r.error).toBe("daemon_cannot_create_nodes");
+    expect(r.blocked_reason).toBe("anet_bin_unknown");
+  });
+
+  test("can_create_nodes=true daemon dispatches normally (gate must not false-positive on healthy daemon)", async () => {
+    seed({
+      role: "host_supervisor",
+      daemon_capabilities: {
+        runtimes_supported: ["claude-agent-sdk"],
+        can_create_nodes: true,
+      },
+    });
+    const r = await call(tools().create_node, { daemon_node_id: DAEMON_ID, node_spec: spec });
+    // Doesn't matter whether downstream dispatch succeeds — the point
+    // is that THIS gate did not reject. If reject, r.error would be
+    // "daemon_cannot_create_nodes".
+    expect(r.error).not.toBe("daemon_cannot_create_nodes");
+  });
+
+  test("🔴 pre-#1371 compat: snapshot with NO can_create_nodes field passes through (must NOT fail-closed)", async () => {
+    // preview.10 through preview.55 daemons never report this field.
+    // If we fail-closed on undefined we silently break every in-flight
+    // daemon that hasn't upgraded. Guard is "known-blocked", not
+    // "unknown-treated-as-blocked".
+    seed({
+      role: "host_supervisor",
+      daemon_capabilities: {
+        runtimes_supported: ["claude-agent-sdk"],
+        // no can_create_nodes at all
+      },
+    });
+    const r = await call(tools().create_node, { daemon_node_id: DAEMON_ID, node_spec: spec });
+    expect(r.error).not.toBe("daemon_cannot_create_nodes");
+  });
+
+  test("pre-#1371 compat: snapshot with no daemon_capabilities at all → passes through", async () => {
+    seed({ role: "host_supervisor" });
+    const r = await call(tools().create_node, { daemon_node_id: DAEMON_ID, node_spec: spec });
+    expect(r.error).not.toBe("daemon_cannot_create_nodes");
+  });
+
+  test("pre-#1371 compat: NULL config_snapshot passes through (very old daemon shape)", async () => {
+    seed(null);
+    const r = await call(tools().create_node, { daemon_node_id: DAEMON_ID, node_spec: spec });
+    expect(r.error).not.toBe("daemon_cannot_create_nodes");
+  });
+
+  test("malformed config_snapshot: permissive fallthrough (mirrors runtimes_supported catch), no crash", async () => {
+    seed("{not json");
+    const r = await call(tools().create_node, { daemon_node_id: DAEMON_ID, node_spec: spec });
+    expect(r.error).not.toBe("daemon_cannot_create_nodes");
+  });
+});
+
+/* #1545 —— 拒绝的时候还要说**那个判断是什么时候做出来的**。
+ *
+ * 为什么这一格落在拒绝路径上最要紧:用户被挡住的这一刻,正好在决定"去修哪台机器"。
+ *   刚测的 blocked      ⇒ 去那台机器按 reason 修
+ *   三周前测的 blocked  ⇒ 那台 daemon 是**开机算一次就永久缓存**的旧版本
+ *                        (agent-node ≤ 2.5.0-preview.54),它可能早就好了 —— 先重启/升级它
+ * 只给 `blocked_reason` 的话,这两种情况给用户的字**一模一样**。 */
+describe("#1545 拒绝文案带上「这个判断是多久以前做的」", () => {
+  const blockedSnap = (extra: object) => ({
+    role: "host_supervisor",
+    daemon_capabilities: {
+      can_create_nodes: false,
+      create_nodes_blocked_reason: "anet_bin_permission",
+      ...extra,
+    },
+  });
+
+  /* 🔴 报出去的是**到现在为止**的绝对年龄,不是 daemon 报的那个原始时长。
+   * daemon 报的是「测完到**这份 report 发出**」;直接透传会漏掉心跳那一段,
+   * **朝更新鲜的方向低报** —— 正是这一格要防的方向。
+   *   绝对年龄 = (now − last_seen_at) + daemon 报的时长 */
+  test("daemon 报了年龄 + 有心跳时间 → 补上心跳那一段,标记 known", async () => {
+    seed(blockedSnap({ create_capability_observed_ms_ago: 12_345 }));
+    seedHeartbeat(60_000);            // 上次心跳在 60s 前
+    const r = await call(tools().create_node, { daemon_node_id: DAEMON_ID, node_spec: spec });
+    expect(r.ok).toBe(false);
+    expect(r.blocked_reason).toBe("anet_bin_permission");
+    expect(r.capability_age).toBe("known");
+    // 允许 ±5s 的执行抖动;关键是它**明显大于** 12,345,即心跳那一段确实被加上了
+    const got = r.capability_observed_ms_ago as number;
+    expect(got).toBeGreaterThanOrEqual(12_345 + 60_000 - 5_000);
+    expect(got).toBeLessThanOrEqual(12_345 + 60_000 + 5_000);
+    // 🔴 反向锚:如果实现是"原样透传",这个数会恰好等于 12,345
+    expect(got).not.toBe(12_345);
+  });
+
+  /* 🔴 第三种 unknown,不能和「旧 daemon 没报」塌成同一个:
+   * 它报了,但 hub 这边没有可用的心跳时间 ⇒ **补不出绝对年龄**,而不是"年龄是 0"。 */
+  test("报了年龄但 hub 没有心跳时间 → unknown_no_heartbeat_time(不是 legacy,也不是 0)", async () => {
+    seed(blockedSnap({ create_capability_observed_ms_ago: 12_345 }));
+    // 故意不 seedHeartbeat
+    const r = await call(tools().create_node, { daemon_node_id: DAEMON_ID, node_spec: spec });
+    expect(r.capability_age).toBe("unknown_no_heartbeat_time");
+    expect(r.capability_observed_ms_ago).toBeNull();
+  });
+
+  /* 🔴 本组的重点:年龄未知时给的是**显式 null + 说明为什么的枚举**,
+   * 而不是省略这两个键。省略读起来像"没有这个问题",而调用方(常常是个 agent)
+   * 会默认它是刚测的 —— 那是朝「更可信」方向撒谎,比不给更糟。 */
+  test("🔴 daemon 没报年龄 → 两个键都在:null + unknown_legacy_daemon(不是省略)", async () => {
+    seed(blockedSnap({}));
+    seedHeartbeat(1_000);   // 心跳齐全 ⇒ 唯一缺的就是 daemon 没报那一格
+    const r = await call(tools().create_node, { daemon_node_id: DAEMON_ID, node_spec: spec });
+    expect(r.ok).toBe(false);
+    expect("capability_observed_ms_ago" in r).toBe(true);
+    expect(r.capability_observed_ms_ago).toBeNull();
+    expect(r.capability_age).toBe("unknown_legacy_daemon");
+    // 🔴 绝不能渲染成"刚测的"
+    expect(r.capability_observed_ms_ago).not.toBe(0);
+  });
+
+  test.each([
+    ["负数", -1],
+    ["超过一年", 400 * 24 * 60 * 60 * 1000],
+    ["非数字", "12"],
+    ["null", null],
+  ])("坏值当作没报(不 clamp 成一个看起来正常的年龄):%s", async (_l, v) => {
+    seed(blockedSnap({ create_capability_observed_ms_ago: v }));
+    // 🔴 必须给心跳:否则这几条会因为"没有心跳时间"而变成
+    //    unknown_no_heartbeat_time,**坏值那一维根本没被测到**
+    //    (夹具只该在被测的那一维上不合法)。
+    seedHeartbeat(1_000);
+    const r = await call(tools().create_node, { daemon_node_id: DAEMON_ID, node_spec: spec });
+    expect(r.capability_age).toBe("unknown_legacy_daemon");
+    expect(r.capability_observed_ms_ago).toBeNull();
+    // 分母自证:拒绝本身仍然发生了,不是因为整条路径没走到才"没有年龄"
+    expect(r.error).toBe("daemon_cannot_create_nodes");
+  });
+
+  test("边界:恰好一年被接受 ⇒ 上面那些不是因为恒判 unknown 才绿", async () => {
+    const oneYear = 365 * 24 * 60 * 60 * 1000;
+    seed(blockedSnap({ create_capability_observed_ms_ago: oneYear }));
+    seedHeartbeat(1_000);
+    const r = await call(tools().create_node, { daemon_node_id: DAEMON_ID, node_spec: spec });
+    expect(r.capability_age).toBe("known");
+    expect(r.capability_observed_ms_ago as number).toBeGreaterThanOrEqual(oneYear);
+  });
+
+  /* 🔴 hub 出**代码**,不出渲染好的修法命令 —— 那张 code → 命令 的表只有一个作者
+   * (@sleep2agi/agent-network 的 daemon-capability-display)。抄一份到 hub,
+   * 两份就会各自漂移,而「hub 教的修法」和「CLI 教的修法」不一致时没有任何东西会红。 */
+  test("🔴 hub 不返回具体修法命令,只把人指到那唯一一份", async () => {
+    seed(blockedSnap({ create_capability_observed_ms_ago: 0 }));
+    seedHeartbeat(1_000);
+    const r = await call(tools().create_node, { daemon_node_id: DAEMON_ID, node_spec: spec });
+    const blob = JSON.stringify(r);
+    for (const leaked of ["chmod go-w", "npm i -g", "readlink -f", "/etc/anet-daemon"]) {
+      expect(blob).not.toContain(leaked);
+    }
+    expect(String(r.remediation_hint)).toContain("anet daemon list");
+    // 正控:这四条不是恒真 —— 它们确实是那份表里的真实内容
+    expect("该二进制 group/other 可写。一行修:chmod go-w \"$(command -v anet)\"").toContain("chmod go-w");
+  });
+
+  test("健康 daemon 不受影响:正常派发,不带这两个键", async () => {
+    seed({ role: "host_supervisor", daemon_capabilities: { can_create_nodes: true } });
+    const r = await call(tools().create_node, { daemon_node_id: DAEMON_ID, node_spec: spec });
+    expect(r.ok).not.toBe(false);
+    expect("capability_age" in r).toBe(false);
+  });
+});

package/src/db.ts

@@ -63,7 +63,7 @@ db.exec(`
   -- 部分索引 acked=0 若「近期」读含已 ack 行则用不上)。P1 只落表。
   CREATE TABLE IF NOT EXISTS user_inbox (
     message_id    TEXT PRIMARY KEY,
-    network_id    TEXT,
+    network_id    TEXT NOT NULL,
     user_id       TEXT NOT NULL,
     from_session  TEXT,
     kind          TEXT NOT NULL DEFAULT 'info',
@@ -681,6 +681,81 @@ function migrateNodeCreateRequestsModelNullable() {
   });
 }
 migrateNodeCreateRequestsModelNullable();
+
+// #1493 — 把 user_inbox.network_id 升到 **schema 级 NOT NULL**(belt-and-suspenders,
+// 叠在 send_desktop_message 的代码级三闸之上:canWrite / `!effectiveNetId` return /
+// getUserNetworkRole,tools.ts)。「不产生 network_id=NULL 孤儿」现是代码级保证 + #1492
+// 三道测试;schema 级约束让**任何**未来绕过那三闸写 NULL 的回归在 INSERT 处直接被
+// 数据库拒(而非产出 scoped-unreadable 的静默孤儿)。
+//
+// SQLite 不能 `ALTER COLUMN ADD NOT NULL` → 建新表→copy→drop→rename(与本文件既有
+// recreate-table 迁移同范式);新库的 CREATE 已直接 NOT NULL,迁移检测到就跳过(幂等)。
+function migrateUserInboxNetworkIdNotNull() {
+  if (db.dialect === "postgres") {
+    try { db.exec("ALTER TABLE user_inbox ALTER COLUMN network_id SET NOT NULL"); } catch {}
+    return;
+  }
+  let createSql: string | undefined;
+  try {
+    createSql = db.get<{ sql: string }>("SELECT sql FROM sqlite_master WHERE type = 'table' AND name = 'user_inbox'")?.sql;
+  } catch { return; }
+  if (!createSql) return;                                       // 表不存在:上面的 CREATE 已建成 NOT NULL 新表
+  if (/network_id\s+TEXT\s+NOT\s+NULL/i.test(createSql)) return; // 已是 NOT NULL → 幂等跳过
+
+  // 🔴 迁移前守卫:存量若有 network_id IS NULL 行(按代码级三闸理应 0)→ **大声 warn +
+  // 跳过迁移**(列暂保持可空),hub 照常启动。fail-safe,不 throw(SDK马 review #1516):
+  //   · 这条 NOT NULL 是 belt-and-suspenders,代码级三闸(#1492 三测钉着)已挡新 NULL;
+  //     为一条冗余保险带让**整个舰队的 hub** boot 期停摆,代价与收益不成比例;
+  //   · 生产这个 COUNT 没人量过,而 #1493 第一条原则就是"没量过的数不要动手"——只要生产
+  //     有一行历史 NULL,throw 就把升级变成停机;
+  //   · fail-closed 的真正目的(不盲目清空、不 COALESCE 猜网络)用 warn+skip 一样达到。
+  // ⚠ 校正:本文件既有迁移(migrateNodeCreateRequestsModelNullable 内部 catch{}+return;
+  //   migrateSessionsNetworkAliasUnique 全函数 throw 0 次)都**不**在 boot 期打死 hub——
+  //   "顶层无 try/catch"是真前提,但"所以 throw 符合惯例"是没验的推论(SDK马 量出),撤回。
+  // 孤儿行仍有明确运维信号,三闸仍生效;等人工量过、清干净,下次启动幂等判据自动完成迁移。
+  const nullCount = db.get<{ n: number }>("SELECT COUNT(*) AS n FROM user_inbox WHERE network_id IS NULL")?.n ?? 0;
+  if (nullCount > 0) {
+    console.error(
+      `[migrate #1493] user_inbox 有 ${nullCount} 行 network_id IS NULL —— 与 send_desktop_message ` +
+      `的代码级三闸矛盾,可能存在更早的写路径漏洞;已**跳过** NOT NULL 迁移(列暂保持可空),` +
+      `hub 正常启动。请人工排查这些孤儿行来源、清理后下次启动会自动完成迁移(不要盲目清空)。`,
+    );
+    return;
+  }
+
+  db.transaction(() => {
+    db.exec("DROP TABLE IF EXISTS user_inbox_migrated");
+    db.exec(`
+      CREATE TABLE user_inbox_migrated (
+        message_id    TEXT PRIMARY KEY,
+        network_id    TEXT NOT NULL,
+        user_id       TEXT NOT NULL,
+        from_session  TEXT,
+        kind          TEXT NOT NULL DEFAULT 'info',
+        title         TEXT,
+        content       TEXT NOT NULL,
+        severity      TEXT NOT NULL DEFAULT 'info',
+        meta_json     TEXT,
+        acked         INTEGER NOT NULL DEFAULT 0,
+        created_at    TEXT NOT NULL DEFAULT (datetime('now')),
+        acked_at      TEXT
+      )
+    `);
+    db.exec(`
+      INSERT INTO user_inbox_migrated
+        (message_id, network_id, user_id, from_session, kind, title, content, severity, meta_json, acked, created_at, acked_at)
+      SELECT
+        message_id, network_id, user_id, from_session, kind, title, content, severity, meta_json, acked, created_at, acked_at
+      FROM user_inbox
+    `);
+    db.exec("DROP TABLE user_inbox");
+    db.exec("ALTER TABLE user_inbox_migrated RENAME TO user_inbox");
+  });
+}
+migrateUserInboxNetworkIdNotNull();
+// 迁移会 drop 掉旧表上的 idx_user_inbox_user_acked → 重建(IF NOT EXISTS;未迁移时 no-op)。
+db.exec(`CREATE INDEX IF NOT EXISTS idx_user_inbox_user_acked ON user_inbox(user_id, acked, created_at);`);
+
 db.exec(`
   CREATE INDEX IF NOT EXISTS idx_ncr_daemon_status ON node_create_requests(daemon_node_id, status);
   CREATE INDEX IF NOT EXISTS idx_ncr_network ON node_create_requests(network_id);

package/src/report-status-host-schema.test.ts

@@ -74,3 +74,138 @@ describe("report_status host telemetry schema", () => {
     })).not.toThrow();
   });
 });
+
+/* #1545 —— `create_capability_observed_ms_ago` 的**验证宽度**本身就是判据。
+ *
+ * 这一格是纯诊断信息(那个 can_create_nodes 是多久以前测出来的)。zod 对象里
+ * 任何一个**已知字段**验证失败,整份 report_status 都会被拒 —— 不是丢掉这一格。
+ * 所以给它加 `.int()/.min()/.max()` 等于:一台节点发了个奇怪的诊断数字,
+ * 就在 hub 上整个失联。**#1225/#1498 已经用 host.ip 演过一遍这件事了。**
+ *
+ * 消毒在读取处(`/api/host-supervisors`),不在这里。 */
+describe("#1545 create_capability_observed_ms_ago —— schema 必须收得宽", () => {
+  const caps = () => schemas.report_status.config_snapshot;
+
+  test("前提:抓到的确实是 config_snapshot 形状,且它接受 daemon_capabilities", () => {
+    expect(caps()).toBeTruthy();
+    const parsed = caps().parse({
+      flags: {}, config_update_capable: false, peer_reply_inbox_capable: true,
+      daemon_capabilities: { can_create_nodes: true },
+    });
+    expect(parsed.daemon_capabilities.can_create_nodes).toBe(true);
+  });
+
+  test("正常值:透传", () => {
+    const parsed = caps().parse({
+      flags: {}, config_update_capable: false, peer_reply_inbox_capable: true,
+      daemon_capabilities: { can_create_nodes: true, create_capability_observed_ms_ago: 0 },
+    });
+    expect(parsed.daemon_capabilities.create_capability_observed_ms_ago).toBe(0);
+  });
+
+  test("null 必须被接受(显式「我没测过」和「我不发这一格」都要能表达)", () => {
+    const parsed = caps().parse({
+      flags: {}, config_update_capable: false, peer_reply_inbox_capable: true,
+      daemon_capabilities: { can_create_nodes: true, create_capability_observed_ms_ago: null },
+    });
+    expect(parsed.daemon_capabilities.create_capability_observed_ms_ago).toBeNull();
+  });
+
+  /* 🔴 这四个值**一个都不许让整份 report 被拒**。它们在读取侧会被当成「没报」,
+   *    但那是读取侧的事 —— 这一层的职责只有一个:别把节点踢下线。 */
+  test.each([
+    ["负数", -1],
+    ["非整数", 1234.5678],
+    ["超过一年", 400 * 24 * 60 * 60 * 1000],
+    ["NaN", Number.NaN],
+  ])("怪值不得拒掉整份 report_status:%s", (_label, value) => {
+    expect(() => caps().parse({
+      flags: {}, config_update_capable: false, peer_reply_inbox_capable: true,
+      daemon_capabilities: { can_create_nodes: false,
+        create_nodes_blocked_reason: "anet_bin_source",
+        create_capability_observed_ms_ago: value },
+    })).not.toThrow();
+  });
+
+  /* 🔴 反向锚:上面那组「不抛」如果只是因为**这个 key 压根没被 schema 看见**,
+   * 它们会一样绿。所以这里证明 schema 确实在看它:喂一个非数字,值必须**被改写成
+   * null**(而不是原样穿过去)。原样穿过去 = key 被忽略,那组测试就什么也没测。 */
+  test("反向锚:非数字被改写成 null(不是原样穿过)⇒ 证明 schema 确实在看这个 key", () => {
+    const parsed = caps().parse({
+      flags: {}, config_update_capable: false, peer_reply_inbox_capable: true,
+      daemon_capabilities: { can_create_nodes: true, create_capability_observed_ms_ago: "0" },
+    });
+    expect(parsed.daemon_capabilities.create_capability_observed_ms_ago).toBeNull();
+    expect(parsed.daemon_capabilities.create_capability_observed_ms_ago).not.toBe("0");
+  });
+});
+
+/* 🔴 #1545 —— 把 `host:` 上方那段方框注释里的表**钉住**。
+ *
+ * 那段注释说的是:这份 schema 的子对象一半 strict 一半不 strict,而「往里加一个
+ * hub 还不认识的键」的后果完全相反 —— 非 strict 静默丢弃(节点活着、字段蒸发),
+ * strict 拒掉整份 report(节点当场失联,#1225 的形状)。
+ *
+ * 注释会烂。这组测试的作用**不是禁止改 strict**(两种都各有其位),而是:谁改了,
+ * 谁就会在这里撞红,从而**被迫连注释一起改**。所以断言写成「和注释里那张表一致」,
+ * 而不是「必须是某个值」。 */
+describe("#1545 report_status 子对象的 strict 分布(注释里那张表的可执行副本)", () => {
+  const rs = () => schemas.report_status;
+  const UNKNOWN = { __hub_does_not_know_this_key__: 1 };
+
+  function acceptsUnknownKey(schema: any, base: object): "dropped" | "rejected" {
+    try {
+      const parsed = schema.parse({ ...base, ...UNKNOWN });
+      return "__hub_does_not_know_this_key__" in parsed ? "rejected" /* 不可能:留着说明它根本没被处理 */ : "dropped";
+    } catch { return "rejected"; }
+  }
+
+  const SIDE_THREAD = {
+    supported: true, runtime: "codex", runtimeVersion: "1.0",
+    topology: "native-exact-fork", evidenceRevision: "r1",
+  };
+  const SNAP_BASE = { flags: {}, config_update_capable: false, peer_reply_inbox_capable: true };
+
+  test.each([
+    ["host", () => rs().host, {}, "dropped"],
+    ["process_telemetry", () => rs().process_telemetry, {}, "dropped"],
+    ["config_snapshot", () => rs().config_snapshot, SNAP_BASE, "dropped"],
+    ["external_schedules", () => rs().external_schedules, { schedules: [] }, "rejected"],
+  ])("%s → 未知键被 %s", (_name, get, base, expected) => {
+    expect(acceptsUnknownKey(get(), base)).toBe(expected as string);
+  });
+
+  test("daemon_capabilities(嵌在 config_snapshot 里)→ 未知键被丢弃,不拒整份", () => {
+    const parsed: any = rs().config_snapshot.parse({
+      ...SNAP_BASE,
+      daemon_capabilities: { can_create_nodes: true, ...UNKNOWN },
+    });
+    expect(parsed.daemon_capabilities.can_create_nodes).toBe(true);
+    expect("__hub_does_not_know_this_key__" in parsed.daemon_capabilities).toBe(false);
+  });
+
+  test("side_thread_capability(同一个 config_snapshot 里)→ 未知键拒掉**整份**", () => {
+    expect(() => rs().config_snapshot.parse({
+      ...SNAP_BASE,
+      side_thread_capability: { ...SIDE_THREAD, ...UNKNOWN },
+    })).toThrow();
+    // 反向锚:去掉那个未知键必须过 —— 否则上面那条可能是被别的必填字段拒的,
+    // 和 strict 无关。
+    expect(() => rs().config_snapshot.parse({
+      ...SNAP_BASE, side_thread_capability: SIDE_THREAD,
+    })).not.toThrow();
+  });
+
+  /* 分母自证:上面这张表**两侧都有**。如果哪天全变成 dropped(或全 rejected),
+   * 那段注释描述的「不对称」就不存在了,而这条会先红。 */
+  test("这张表确实是不对称的(两种后果各至少一个)", () => {
+    const results = [
+      acceptsUnknownKey(rs().host, {}),
+      acceptsUnknownKey(rs().process_telemetry, {}),
+      acceptsUnknownKey(rs().config_snapshot, SNAP_BASE),
+      acceptsUnknownKey(rs().external_schedules, { schedules: [] }),
+    ];
+    expect(results.filter(r => r === "dropped").length).toBeGreaterThan(0);
+    expect(results.filter(r => r === "rejected").length).toBeGreaterThan(0);
+  });
+});