fix: route master agent through codex device pool

2026-06-06 12:31:04 +08:00
parent 684b98c5c1
commit 9e81d8a960
7 changed files with 619 additions and 32 deletions
--- a/docs/architecture/ai_handoff_index_cn.md
+++ b/docs/architecture/ai_handoff_index_cn.md
@@ -194,6 +194,7 @@
 - 多用户 / RBAC / Skill / 主 Agent 权限和多设备控制的集中状态、回归矩阵与缺口清单见 `docs/architecture/rbac_skill_regression_matrix_cn.md`
 - `我的 > 主 Agent 提示词 / 记忆` 当前可编辑管理员全局主提示词、用户主提示词、当前对话附加提示词，以及用户通用记忆 / 项目记忆
 - `我的 > AI 账号` 必须可查看和切换 `主 GPT / 备用 GPT / API 容灾`
+- 主 Agent 使用 `Master Codex Node` 时必须优先走授权 Codex 设备池：设备在线且 `Codex App Server / CLI / GUI` 至少一条模型通道在线才可用；首选设备不可用或执行失败会自动切下一台，全部 Codex 设备不可用后才使用用户配置的 API Key；如果两类通道都没有，APP 中提示“当前没有可用的模型渠道”
 - `我的 > 技能` 必须按绑定设备展示 Skill，并支持一键复制调用语句
 - Skill 远程治理第一版已经接通最高管理员后端入口和设备端执行：`GET/POST /api/v1/admin/skills/requests` 可创建和查看 `install / update / uninstall / rollback / version_lock` 请求，local-agent 通过 `claim / complete` 认领执行并回写最新 Skill 清单。当前设备端已增加 source allowlist / trusted source、`checksum / expectedChecksum` sha256 校验、更新 / 卸载 / 回滚前备份和失败恢复；仍未做签名校验和依赖安装沙箱
 - `设备` 页当前只允许出现生产设备，旧演示脏数据不能回流到正式视图
--- a/docs/architecture/api_and_service_inventory_cn.md
+++ b/docs/architecture/api_and_service_inventory_cn.md
@@ -759,8 +759,8 @@
  - `projectId=master-agent` 且 `kind=text` 时，会先返回 `masterReplyState + task`，真实回复随后异步回写到账本
  - Telegram Gateway 当前也复用这条主 Agent 链路：Telegram 私聊文本会写入 `master-agent` 项目，快速回复直接返回，异步任务通过 `externalReplyTarget` 在完成后回推 Telegram
  - 当前主链路优先走 `Master Codex Node`：`task queue -> local-agent -> codex exec -> complete`
-  - 如果当前主控是 `Master Codex Node`，但节点离线或执行立即失败，主 Agent 当前会优先尝试已配置的 `OpenAI API / 阿里百炼 Qwen` 账号，避免聊天直接只剩失败日志
-  - 如本机节点未接通，可切到 `OpenAI API` 或 `阿里百炼 Qwen` 备用账号
+  - 如果当前主控是 `Master Codex Node`，主 Agent 会先使用授权范围内的 Codex 设备池：设备在线且 `Codex App Server / CLI / GUI` 至少一条模型通道在线才可作为模型通道；首选设备不可用会切下一台，执行中失败会把同一任务重排到下一台，全部 Codex 设备不可用后才切到已配置的 API 备用链
+  - 如 Codex 设备池和 API Key 都不可用，主 Agent 会在对话里提示“当前没有可用的模型渠道”，不再暴露内部 master 节点账号话术
  - 群聊项目当前会带上 `collaborationGate`，用于标明当前是否需要先经主 Agent / 用户审批
  - 群聊文本消息当前还会返回 `dispatchPlan / dispatchRecommendation`，用于展示主 Agent 推荐的线程下发方案
  - 如果群里已经有一条待确认推荐，接口会直接返回 `409`，要求先确认或拒绝当前推荐，避免审批消息叠加
--- a/docs/architecture/current_runtime_and_deploy_status_cn.md
+++ b/docs/architecture/current_runtime_and_deploy_status_cn.md
@@ -188,7 +188,7 @@ cd /Users/kris/code/boss
 - 当前 `我的 > 主 Agent 提示词 / 记忆` 页面已接通：管理员全局主提示词只读展示、用户主提示词、当前对话附加提示词，以及用户通用记忆 / 跨项目项目记忆都可以在 Web 端查看和编辑；当前对话设置按登录账号隔离，管理员全局主提示词不可覆盖
 - 当前 Web 端 `master-agent` 会话页右上角也已补齐微信式三点菜单，支持直接进入 `提示词 / 模型 / 推理强度 / 记忆 / 刷新`
 - 当前 `approval_required` 群聊在 Web 端已统一用单一状态快照驱动：如果存在新的待确认推荐，会自动折叠旧的拒绝态；如果上次推荐已拒绝，会明确展示“重新生成新的推荐”的恢复入口
- 当前如果主控身份还是 `Master Codex Node`，但该节点离线或执行立即失败，主 Agent 会优先尝试已配置的 `OpenAI API / 阿里百炼 Qwen` 备用账号，不再把失败日志直接原样回给用户
+- 当前如果主控身份是 `Master Codex Node`，主 Agent 会先按授权范围构建 Codex 设备池：设备必须在线，且 `Codex App Server / CLI / GUI` 至少一条模型通道在线，`codexAppServer.metadata.accountSummary.signedIn=false` 会被视为不可用；首选设备不可用时会自动切到下一台可用 Codex 设备，执行中失败也会先重排到下一台设备，全部 Codex 设备不可用时才尝试已配置的 API 备用链；如果 Codex 设备池和 API Key 都不可用，APP 会提示“当前没有可用的模型渠道”
 - 当前原生 Android 的聊天发送已收短客户端等待窗口；`master-agent` 单聊依赖服务端快速入队和消息流里的“主 Agent 思考中 / 回复超时 / 重试等待”状态，不再要求客户端长时间同步阻塞
 - 当前设备导入主链也已补上第一轮后端闭环：`heartbeat` 可上报真实项目候选，服务端会生成 `deviceImportDraft`；用户可提交勾选结果、生成导入决议，再把选中的线程真正落成聊天窗口
 - Web 与原生 Android 当前都已补上“新设备导入草稿 -> 勾选 -> 决议预览 -> 应用导入”的前台流程；已绑定生产设备继续保留 heartbeat 自动导入主链
--- a/src/lib/boss-data.ts
+++ b/src/lib/boss-data.ts
@@ -1333,6 +1333,7 @@ export interface MasterAgentTask {
  deviceId: string;
  accountId?: string;
  accountLabel?: string;
+  modelChannelAttemptedDeviceIds?: string[];
  attachmentId?: string;
  attachmentFileName?: string;
  attachmentDownloadToken?: string;
@@ -9915,6 +9916,123 @@ function isTerminalMasterAgentTaskStatus(status: MasterAgentTaskStatus) {
  return status === "completed" || status === "failed" || status === "timed_out" || status === "canceled";
 }

+function codexModelChannelAccountSummarySignedIn(device: Device) {
+  const metadata = device.capabilities?.codexAppServer?.metadata;
+  const accountSummary =
+    metadata && typeof metadata === "object"
+      ? (metadata.accountSummary as { signedIn?: unknown } | undefined)
+      : undefined;
+  return accountSummary?.signedIn !== false;
+}
+
+function hasUsableCodexModelChannelInState(device: Device | undefined) {
+  if (!device || device.status !== "online" || isDeviceRevoked(device)) {
+    return false;
+  }
+  const capabilities = device.capabilities;
+  const hasCodexTransport = Boolean(
+    capabilities?.codexAppServer?.connected ||
+      capabilities?.cli?.connected ||
+      capabilities?.gui?.connected,
+  );
+  return hasCodexTransport && codexModelChannelAccountSummarySignedIn(device);
+}
+
+function isMasterCodexNodeTask(task: MasterAgentTask, state: BossState) {
+  if (task.taskType !== "conversation_reply") {
+    return false;
+  }
+  if (task.accountId?.startsWith("master-codex-device-")) {
+    return true;
+  }
+  const account = task.accountId
+    ? state.aiAccounts.find((item) => item.accountId === task.accountId)
+    : undefined;
+  return account?.provider === "master_codex_node";
+}
+
+function isTaskAuthorizedForDevice(task: MasterAgentTask, deviceId: string) {
+  return (
+    !task.authorizedDeviceIds ||
+    task.authorizedDeviceIds.length === 0 ||
+    task.authorizedDeviceIds.includes(deviceId)
+  );
+}
+
+function sortMasterCodexFailoverDevices(left: Device, right: Device) {
+  return (right.lastSeenAt ?? "").localeCompare(left.lastSeenAt ?? "");
+}
+
+function findNextMasterCodexNodeCandidateForFailedTask(
+  state: BossState,
+  task: MasterAgentTask,
+  failedDeviceId: string,
+) {
+  const excludedDeviceIds = new Set([failedDeviceId, ...(task.modelChannelAttemptedDeviceIds ?? [])]);
+  const seenDeviceIds = new Set<string>();
+
+  const configuredCandidates = state.aiAccounts
+    .filter((account) => {
+      const deviceId = account.nodeId?.trim();
+      if (
+        !account.enabled ||
+        account.provider !== "master_codex_node" ||
+        (account.status !== "ready" && account.status !== "degraded") ||
+        !deviceId ||
+        excludedDeviceIds.has(deviceId) ||
+        !isTaskAuthorizedForDevice(task, deviceId)
+      ) {
+        return false;
+      }
+      const device = state.devices.find((item) => item.id === deviceId);
+      return hasUsableCodexModelChannelInState(device);
+    })
+    .sort((left, right) => {
+      if (left.isActive !== right.isActive) {
+        return left.isActive ? -1 : 1;
+      }
+      return (right.updatedAt ?? "").localeCompare(left.updatedAt ?? "");
+    });
+
+  const configured = configuredCandidates[0];
+  if (configured?.nodeId) {
+    return {
+      deviceId: configured.nodeId,
+      accountId: configured.accountId,
+      accountLabel: configured.label,
+    };
+  }
+
+  for (const account of configuredCandidates) {
+    if (account.nodeId) {
+      seenDeviceIds.add(account.nodeId);
+    }
+  }
+
+  const device = state.devices
+    .filter((item) => {
+      if (
+        excludedDeviceIds.has(item.id) ||
+        seenDeviceIds.has(item.id) ||
+        !isTaskAuthorizedForDevice(task, item.id)
+      ) {
+        return false;
+      }
+      return hasUsableCodexModelChannelInState(item);
+    })
+    .sort(sortMasterCodexFailoverDevices)[0];
+
+  if (!device) {
+    return null;
+  }
+
+  return {
+    deviceId: device.id,
+    accountId: `master-codex-device-${device.id}`,
+    accountLabel: `Codex · ${device.name || device.id}`,
+  };
+}
+
 function masterAgentTaskLeaseMs(task: MasterAgentTask) {
  return task.taskType === "conversation_reply"
    ? MASTER_AGENT_TASK_CONVERSATION_LEASE_MS
@@ -10376,6 +10494,46 @@ export async function completeMasterAgentTask(payload: {
        dispatchExecution: undefined,
      };
    }
+    if (payload.status === "failed" && isMasterCodexNodeTask(task, state)) {
+      const failover = findNextMasterCodexNodeCandidateForFailedTask(state, task, payload.deviceId);
+      if (failover) {
+        const failedAt = nowIso();
+        const failedAccount = task.accountId
+          ? state.aiAccounts.find((item) => item.accountId === task.accountId)
+          : undefined;
+        if (failedAccount?.provider === "master_codex_node") {
+          failedAccount.status = "degraded";
+          failedAccount.lastError = payload.errorMessage?.trim() || "MASTER_CODEX_NODE_EXEC_FAILED";
+          failedAccount.lastValidatedAt = failedAt;
+          failedAccount.updatedAt = failedAt;
+        }
+        task.status = "queued";
+        task.deviceId = failover.deviceId;
+        task.accountId = failover.accountId;
+        task.accountLabel = failover.accountLabel;
+        task.modelChannelAttemptedDeviceIds = [
+          ...new Set([...(task.modelChannelAttemptedDeviceIds ?? []), payload.deviceId]),
+        ];
+        task.completedAt = undefined;
+        task.canceledAt = undefined;
+        task.canceledBy = undefined;
+        task.cancelReason = undefined;
+        task.leaseExpiresAt = undefined;
+        task.claimedAt = undefined;
+        task.lastClaimedAt = undefined;
+        task.lastErrorKind = "model_channel_failover";
+        task.errorMessage = payload.errorMessage?.trim() || "MASTER_CODEX_NODE_EXEC_FAILED";
+        task.replyBody = undefined;
+        task.requestId = undefined;
+        upsertTaskExecutionProgressMessageInState(state, task, "queued", payload.executionProgress);
+        return {
+          ...task,
+          dispatchPlan: undefined,
+          dispatchExecution: undefined,
+          dialogGuardIntervention: undefined,
+        };
+      }
+    }
    task.status = payload.status;
    task.completedAt = nowIso();
    task.leaseExpiresAt = undefined;
--- a/src/lib/boss-master-agent.ts
+++ b/src/lib/boss-master-agent.ts
@@ -25,6 +25,7 @@ import type {
  AiAccount,
  AiProvider,
  AuthRole,
+  Device,
  BossPermission,
  DispatchPlanTarget,
  ExternalReplyTarget,
@@ -1740,9 +1741,36 @@ function isUsableMasterNodeAccount(account: AiAccount) {
  );
 }

+function codexAppServerAccountSummarySignedIn(device: Device) {
+  const metadata = device.capabilities?.codexAppServer?.metadata;
+  const accountSummary =
+    metadata && typeof metadata === "object"
+      ? (metadata.accountSummary as { signedIn?: unknown } | undefined)
+      : undefined;
+  return accountSummary?.signedIn !== false;
+}
+
+function hasUsableCodexModelChannel(device: Device | undefined) {
+  if (!device || device.status !== "online") {
+    return false;
+  }
+  const capabilities = device.capabilities;
+  const hasCodexTransport = Boolean(
+    capabilities?.codexAppServer?.connected ||
+      capabilities?.cli?.connected ||
+      capabilities?.gui?.connected,
+  );
+  return hasCodexTransport && codexAppServerAccountSummarySignedIn(device);
+}
+
+function isAuthorizedDeviceId(deviceId: string, authorizedDeviceIds?: string[]) {
+  return !authorizedDeviceIds || authorizedDeviceIds.length === 0 || authorizedDeviceIds.includes(deviceId);
+}
+
 function isOnlineMasterNodeAccount(
  state: Awaited<ReturnType<typeof readState>>,
  account: AiAccount,
+  authorizedDeviceIds?: string[],
 ) {
  if (!isUsableMasterNodeAccount(account)) {
    return false;
@@ -1751,8 +1779,11 @@ function isOnlineMasterNodeAccount(
  if (!deviceId) {
    return false;
  }
+  if (!isAuthorizedDeviceId(deviceId, authorizedDeviceIds)) {
+    return false;
+  }
  const device = state.devices.find((item) => item.id === deviceId);
-  return Boolean(device && device.status === "online");
+  return hasUsableCodexModelChannel(device);
 }

 function sortSelectableAccounts(left: AiAccount, right: AiAccount) {
@@ -1790,19 +1821,106 @@ function supportsResponsesReasoning(provider: ApiCompatibleProvider) {
  return provider === "openai_api" || provider === "hyzq_api";
 }

+function buildAutoDiscoveredMasterNodeAccount(device: Device): AiAccount {
+  const now = device.lastSeenAt || new Date(0).toISOString();
+  return {
+    accountId: `master-codex-device-${device.id}`,
+    label: `Codex · ${device.name || device.id}`,
+    role: "backup",
+    provider: "master_codex_node",
+    displayName: `${device.name || device.id} 上的 Codex`,
+    nodeId: device.id,
+    nodeLabel: device.name || device.id,
+    model: "Codex",
+    enabled: true,
+    isActive: false,
+    status: "ready",
+    loginStatusNote: "由绑定设备 heartbeat 自动发现的 Codex 模型通道。",
+    lastValidatedAt: device.lastSeenAt,
+    lastUsedAt: device.lastSeenAt,
+    createdAt: now,
+    updatedAt: now,
+  };
+}
+
+function sortAutoDiscoveredCodexDevices(left: Device, right: Device) {
+  return (right.lastSeenAt ?? "").localeCompare(left.lastSeenAt ?? "");
+}
+
+function resolveMasterCodexNodeExecutionCandidates(params: {
+  state: Awaited<ReturnType<typeof readState>>;
+  runtimeAccount: AiAccount;
+  authorizedDeviceIds?: string[];
+  excludeDeviceIds?: string[];
+}) {
+  const seenAccountIds = new Set<string>();
+  const seenDeviceIds = new Set<string>();
+  const excludedDeviceIds = new Set(params.excludeDeviceIds ?? []);
+  const candidates: AiAccount[] = [];
+
+  const pushConfiguredAccount = (account: AiAccount | undefined) => {
+    if (!account || seenAccountIds.has(account.accountId)) {
+      return;
+    }
+    const deviceId = account.nodeId?.trim();
+    if (!deviceId || excludedDeviceIds.has(deviceId)) {
+      return;
+    }
+    if (!isOnlineMasterNodeAccount(params.state, account, params.authorizedDeviceIds)) {
+      return;
+    }
+    seenAccountIds.add(account.accountId);
+    seenDeviceIds.add(deviceId);
+    candidates.push(account);
+  };
+
+  if (params.runtimeAccount.provider === "master_codex_node") {
+    pushConfiguredAccount(params.runtimeAccount);
+  }
+
+  for (const account of params.state.aiAccounts
+    .filter((item) => item.accountId !== params.runtimeAccount.accountId)
+    .sort(sortSelectableAccounts)) {
+    pushConfiguredAccount(account);
+  }
+
+  for (const device of params.state.devices
+    .filter((item) => {
+      if (excludedDeviceIds.has(item.id) || seenDeviceIds.has(item.id)) {
+        return false;
+      }
+      if (!isAuthorizedDeviceId(item.id, params.authorizedDeviceIds)) {
+        return false;
+      }
+      return hasUsableCodexModelChannel(item);
+    })
+    .sort(sortAutoDiscoveredCodexDevices)) {
+    const account = buildAutoDiscoveredMasterNodeAccount(device);
+    if (seenAccountIds.has(account.accountId)) {
+      continue;
+    }
+    seenAccountIds.add(account.accountId);
+    seenDeviceIds.add(device.id);
+    candidates.push(account);
+  }
+
+  return candidates;
+}
+
+function isPersistedAiAccount(
+  state: Awaited<ReturnType<typeof readState>>,
+  account: AiAccount,
+) {
+  return state.aiAccounts.some((item) => item.accountId === account.accountId);
+}
+
 async function resolveAccountForSelectedBackend(
  selectedBackendProvider: AiProvider,
  runtimeAccount: AiAccount,
 ) {
  if (selectedBackendProvider === "master_codex_node") {
    const state = await readState();
-    if (isOnlineMasterNodeAccount(state, runtimeAccount)) {
-      return runtimeAccount;
-    }
-
-    return state.aiAccounts
-      .filter((account) => isOnlineMasterNodeAccount(state, account))
-      .sort(sortSelectableAccounts)[0];
+    return resolveMasterCodexNodeExecutionCandidates({ state, runtimeAccount })[0] ?? null;
  }

  if (isApiCompatibleProvider(selectedBackendProvider)) {
@@ -1993,13 +2111,16 @@ export async function tryBuildLocalMasterAgentFastReply(params: {
 async function resolveMasterNodeExecutionCandidate(params: {
  backendChoices: Array<{ backendId: string; provider?: AiProvider }>;
  runtimeAccount: AiAccount;
+  candidates?: AiAccount[];
 }) {
  const wantsMasterNode = params.backendChoices.some((backend) => backend.backendId === "master-codex-node");
  if (!wantsMasterNode) {
    return null;
  }

-  const account = await resolveAccountForSelectedBackend("master_codex_node", params.runtimeAccount);
+  const account =
+    params.candidates?.[0] ??
+    (await resolveAccountForSelectedBackend("master_codex_node", params.runtimeAccount));
  return account && account.provider === "master_codex_node" ? account : null;
 }

@@ -4160,10 +4281,13 @@ export async function replyToMasterAgentUserMessage(params: {
    return { ok: false as const, reason: "FORBIDDEN" };
  }
  const replyProject = authorizedState.projects.find((project) => project.id === replyProjectId);
-  const primaryDeviceId = runtime.account.nodeId || state.user.boundDeviceId || "mac-studio";
-  const primaryDevice = state.devices.find((device) => device.id === primaryDeviceId);
+  const masterNodeCandidates = resolveMasterCodexNodeExecutionCandidates({
+    state,
+    runtimeAccount: runtime.account,
+    authorizedDeviceIds: authorizedScope.authorizedDeviceIds,
+  });
  const primaryBackendStatus =
-    runtime.account.provider === "master_codex_node" && (!primaryDevice || primaryDevice.status !== "online")
+    runtime.account.provider === "master_codex_node" && masterNodeCandidates.length === 0
      ? "degraded"
      : runtime.account.status;
  const clawSelectionState = await getClawBackendSelectionState();
@@ -4176,7 +4300,11 @@ export async function replyToMasterAgentUserMessage(params: {
      .filter((account) => account.accountId !== runtime.account.accountId)
      .map((account) => ({
        provider: account.provider,
-        status: account.status,
+        status:
+          account.provider === "master_codex_node" &&
+          !isOnlineMasterNodeAccount(state, account, authorizedScope.authorizedDeviceIds)
+            ? "degraded"
+            : account.status,
      })),
    requestKind: "master_agent_reply" as const,
    requestedBackendId: executionConfig.agentControls?.backendOverride,
@@ -4194,6 +4322,7 @@ export async function replyToMasterAgentUserMessage(params: {
  const selectedMasterAccount = await resolveMasterNodeExecutionCandidate({
    backendChoices,
    runtimeAccount: runtime.account,
+    candidates: masterNodeCandidates,
  });
  const apiExecutionCandidates = await buildApiExecutionCandidates({
    backendChoices,
@@ -4333,8 +4462,8 @@ export async function replyToMasterAgentUserMessage(params: {
    if (!selectedMasterAccount) {
      await appendMasterAgentSystemReply(
        [
-          `当前主控身份是 ${runtime.summary.roleLabel}，目标后端是 Master Codex Node，但当前没有可用的 master 节点账号。`,
-          "请先把可用的 Master Codex Node 重新接回，再重试。",
+          "当前没有可用的模型渠道。",
+          "请先接回一台已经登录 Codex / ChatGPT 的电脑，或到“我的 > AI 账号”配置可用 API Key 后再重试。",
        ].join(""),
        `主 Agent · ${runtime.summary.roleLabel}`,
        replyProjectId,
@@ -4351,15 +4480,27 @@ export async function replyToMasterAgentUserMessage(params: {
      state.user.boundCodexNodeLabel ||
      deviceId;

-    if (!boundDevice || boundDevice.status !== "online") {
-      await updateAiAccountHealth({
-        accountId: selectedMasterAccount.accountId,
-        status: "degraded",
-        lastError: !boundDevice ? "MASTER_CODEX_NODE_DEVICE_NOT_FOUND" : "MASTER_CODEX_NODE_DEVICE_OFFLINE",
-        lastValidatedAt: new Date().toISOString(),
-      });
+    if (!hasUsableCodexModelChannel(boundDevice)) {
+      const lastError = !boundDevice
+        ? "MASTER_CODEX_NODE_DEVICE_NOT_FOUND"
+        : boundDevice.status !== "online"
+          ? "MASTER_CODEX_NODE_DEVICE_OFFLINE"
+          : "MASTER_CODEX_NODE_CHANNEL_UNAVAILABLE";
+      if (isPersistedAiAccount(state, selectedMasterAccount)) {
+        await updateAiAccountHealth({
+          accountId: selectedMasterAccount.accountId,
+          status: "degraded",
+          lastError,
+          lastValidatedAt: new Date().toISOString(),
+        });
+      }
+      const unavailableReason = !boundDevice
+        ? "未找到"
+        : boundDevice.status !== "online"
+          ? "不在线"
+          : "没有检测到可用的 Codex 模型通道";
      await appendMasterAgentSystemReply(
-        `主 GPT 不在手机里直接登录。当前绑定设备 ${boundNodeLabel}${boundDevice ? " 不在线" : " 未找到"}，主 Agent 暂时无法通过这台设备对话。请先在该设备上登录 Codex / ChatGPT Plus，并确保 local-agent 在线后再重试。`,
+        `主 GPT 不在手机里直接登录。当前绑定设备 ${boundNodeLabel} ${unavailableReason}，主 Agent 暂时无法通过这台设备对话。请先在该设备上登录 Codex / ChatGPT Plus，并确保 boss-agent 在线后再重试。`,
        `主 Agent · ${selectedMasterAccount.label || runtime.summary.roleLabel}`,
        replyProjectId,
        params.requestedByAccount,
@@ -4367,12 +4508,14 @@ export async function replyToMasterAgentUserMessage(params: {
      return { ok: false as const, reason: "MASTER_NODE_OFFLINE" };
    }

-    await updateAiAccountHealth({
-      accountId: selectedMasterAccount.accountId,
-      status: "ready",
-      lastValidatedAt: new Date().toISOString(),
-      lastUsedAt: new Date().toISOString(),
-    });
+    if (isPersistedAiAccount(state, selectedMasterAccount)) {
+      await updateAiAccountHealth({
+        accountId: selectedMasterAccount.accountId,
+        status: "ready",
+        lastValidatedAt: new Date().toISOString(),
+        lastUsedAt: new Date().toISOString(),
+      });
+    }

    if (
      controlIntent.intentCategory === "browser_control" || controlIntent.intentCategory === "desktop_control"
--- a/tests/master-agent-message-queue.test.ts
+++ b/tests/master-agent-message-queue.test.ts
@@ -10,6 +10,7 @@ let POST: (typeof import("../src/app/api/v1/projects/[projectId]/messages/route"
 let saveAiAccount: (typeof import("../src/lib/boss-data"))["saveAiAccount"];
 let updateProjectAgentControls: (typeof import("../src/lib/boss-data"))["updateProjectAgentControls"];
 let updateAiAccountHealth: (typeof import("../src/lib/boss-data"))["updateAiAccountHealth"];
+let updateDevice: (typeof import("../src/lib/boss-data"))["updateDevice"];
 let readState: (typeof import("../src/lib/boss-data"))["readState"];
 let createAuthSession: (typeof import("../src/lib/boss-data"))["createAuthSession"];
 let appendProjectMessages: (typeof import("../src/lib/boss-data"))["appendProjectMessages"];
@@ -34,6 +35,7 @@ async function setup() {
  saveAiAccount = data.saveAiAccount;
  updateProjectAgentControls = data.updateProjectAgentControls;
  updateAiAccountHealth = data.updateAiAccountHealth;
+  updateDevice = data.updateDevice;
  readState = data.readState;
  createAuthSession = data.createAuthSession;
  appendProjectMessages = data.appendProjectMessages;
@@ -844,6 +846,14 @@ test("POST /api/v1/projects/master-agent/messages 在主节点在线时复杂快
 });

 test("master-agent enqueue 在主节点离线时会自动切到 OpenAI 后台队列而不是挂到本机设备队列", async () => {
+  await updateDevice("mac-studio", {
+    capabilities: {
+      gui: { connected: false },
+      cli: { connected: false },
+      codexAppServer: { connected: false },
+    },
+  });
+
  await saveAiAccount({
    accountId: "master-codex-primary-offline",
    label: "主 GPT",
--- a/tests/master-agent-openai-fallback.test.ts
+++ b/tests/master-agent-openai-fallback.test.ts
@@ -10,6 +10,9 @@ let saveAiAccount: (typeof import("../src/lib/boss-data"))["saveAiAccount"];
 let readState: (typeof import("../src/lib/boss-data"))["readState"];
 let updateAiAccountHealth: (typeof import("../src/lib/boss-data"))["updateAiAccountHealth"];
 let updateProjectAgentControls: (typeof import("../src/lib/boss-data"))["updateProjectAgentControls"];
+let updateDevice: (typeof import("../src/lib/boss-data"))["updateDevice"];
+let upsertDeviceHeartbeat: (typeof import("../src/lib/boss-data"))["upsertDeviceHeartbeat"];
+let completeMasterAgentTask: (typeof import("../src/lib/boss-data"))["completeMasterAgentTask"];

 async function setup() {
  if (runtimeRoot) return;
@@ -28,6 +31,9 @@ async function setup() {
  readState = data.readState;
  updateAiAccountHealth = data.updateAiAccountHealth;
  updateProjectAgentControls = data.updateProjectAgentControls;
+  updateDevice = data.updateDevice;
+  upsertDeviceHeartbeat = data.upsertDeviceHeartbeat;
+  completeMasterAgentTask = data.completeMasterAgentTask;
 }

 async function waitFor(predicate: () => Promise<boolean>, timeoutMs = 5_000) {
@@ -54,6 +60,14 @@ test.beforeEach(async () => {
 });

 test("replyToMasterAgentUserMessage falls back to a runnable OpenAI API account when the master node is offline", async () => {
+  await updateDevice("mac-studio", {
+    capabilities: {
+      gui: { connected: false },
+      cli: { connected: false },
+      codexAppServer: { connected: false },
+    },
+  });
+
  await saveAiAccount({
    accountId: "master-codex-primary",
    label: "主 GPT",
@@ -253,6 +267,14 @@ test("replyToMasterAgentUserMessage uses active DeepSeek API accounts directly",
 });

 test("replyToMasterAgentUserMessage falls back to a runnable aliyun qwen backup account when the master node is offline", async () => {
+  await updateDevice("mac-studio", {
+    capabilities: {
+      gui: { connected: false },
+      cli: { connected: false },
+      codexAppServer: { connected: false },
+    },
+  });
+
  await saveAiAccount({
    accountId: "master-codex-primary",
    label: "主 GPT",
@@ -508,3 +530,256 @@ test("replyToMasterAgentUserMessage falls back to a ready backup master node acc
  assert.equal(task?.accountId, "master-codex-backup-ready");
  assert.equal(task?.deviceId, "mac-studio");
 });
+
+test("replyToMasterAgentUserMessage routes to another bound Codex device when the active node has no usable Codex channel", async () => {
+  await updateDevice("mac-studio", {
+    status: "online",
+    capabilities: {
+      gui: { connected: false },
+      cli: { connected: false },
+      codexAppServer: { connected: false },
+    },
+  });
+  await upsertDeviceHeartbeat({
+    deviceId: "macbook-air-codex",
+    token: "boss-macbook-air-codex-token",
+    name: "MacBook Air",
+    avatar: "A",
+    account: "krisolo",
+    status: "online",
+    quota5h: 90,
+    quota7d: 90,
+    projects: [],
+    capabilities: {
+      gui: { connected: true },
+      cli: { connected: true },
+      codexAppServer: { connected: true },
+    },
+  });
+
+  await saveAiAccount({
+    accountId: "master-codex-primary-no-channel",
+    label: "主 GPT",
+    role: "primary",
+    provider: "master_codex_node",
+    displayName: "在线但没有 Codex 通道的主节点",
+    nodeId: "mac-studio",
+    nodeLabel: "Mac Studio",
+    model: "gpt-5.4",
+    enabled: true,
+    setActive: true,
+    loginStatusNote: "模拟 boss-agent 未检测到可用 Codex 通道。",
+  });
+
+  const result = await replyToMasterAgentUserMessage({
+    requestMessageId: "msg-master-node-device-pool",
+    requestText: "请使用可用的 Codex 设备池。",
+    requestedBy: "Boss 超级管理员",
+    requestedByAccount: "krisolo",
+    mode: "enqueue",
+  });
+
+  assert.equal(result.ok, true);
+  assert.equal(result.accountId, "master-codex-device-macbook-air-codex");
+  assert.ok(result.taskId, "expected a queued master-agent task");
+
+  const state = await readState();
+  const task = state.masterAgentTasks.find((item) => item.taskId === result.taskId);
+  assert.ok(task, "expected queued task to be written into state");
+  assert.equal(task?.accountId, "master-codex-device-macbook-air-codex");
+  assert.equal(task?.deviceId, "macbook-air-codex");
+});
+
+test("replyToMasterAgentUserMessage falls back to API when no bound Codex model channel is usable", async () => {
+  await updateDevice("mac-studio", {
+    status: "online",
+    capabilities: {
+      gui: { connected: false },
+      cli: { connected: false },
+      codexAppServer: { connected: false },
+    },
+  });
+
+  await saveAiAccount({
+    accountId: "master-codex-primary-no-model-channel",
+    label: "主 GPT",
+    role: "primary",
+    provider: "master_codex_node",
+    displayName: "不可用的 Master Codex Node",
+    nodeId: "mac-studio",
+    nodeLabel: "Mac Studio",
+    model: "gpt-5.4",
+    enabled: true,
+    setActive: true,
+    loginStatusNote: "模拟没有任何可用 Codex 模型通道。",
+  });
+
+  await saveAiAccount({
+    accountId: "openai-api-after-codex-pool",
+    label: "API 备用",
+    role: "api_fallback",
+    provider: "openai_api",
+    displayName: "OpenAI API 备用账号",
+    model: "gpt-5.4-mini",
+    apiKey: "sk-openai-after-codex-pool",
+    enabled: true,
+    setActive: false,
+    loginStatusNote: "Codex 设备池不可用时兜底。",
+  });
+
+  const originalFetch = globalThis.fetch;
+  globalThis.fetch = (async (input) => {
+    if (typeof input === "string" && input === "https://api.openai.com/v1/responses") {
+      return new Response(JSON.stringify({ output_text: "已切到 API 备用链。" }), {
+        status: 200,
+        headers: {
+          "content-type": "application/json",
+          "x-request-id": "req-api-after-codex-pool",
+        },
+      });
+    }
+    throw new Error(`unexpected fetch: ${String(input)}`);
+  }) as typeof fetch;
+
+  try {
+    const result = await replyToMasterAgentUserMessage({
+      requestMessageId: "msg-api-after-codex-pool",
+      requestText: "Codex 设备都不可用时继续回复。",
+      requestedBy: "Boss 超级管理员",
+      requestedByAccount: "krisolo",
+      mode: "enqueue",
+    });
+
+    assert.equal(result.ok, true);
+    assert.equal(result.accountId, "openai-api-after-codex-pool");
+    assert.ok(result.taskId, "expected an API fallback task");
+
+    const state = await readState();
+    const task = state.masterAgentTasks.find((item) => item.taskId === result.taskId);
+    assert.ok(task, "expected queued task to be written into state");
+    assert.equal(task?.accountId, "openai-api-after-codex-pool");
+    assert.equal(task?.deviceId, "master-agent-openai");
+
+    await waitFor(async () => {
+      const nextState = await readState();
+      const completedTask = nextState.masterAgentTasks.find((item) => item.taskId === result.taskId);
+      return completedTask?.status === "completed";
+    });
+  } finally {
+    globalThis.fetch = originalFetch;
+  }
+});
+
+test("failed Master Codex Node task is requeued to the next usable Codex device before surfacing an error", async () => {
+  await updateDevice("mac-studio", {
+    status: "online",
+    capabilities: {
+      gui: { connected: true },
+      cli: { connected: true },
+      codexAppServer: { connected: true },
+    },
+  });
+  await upsertDeviceHeartbeat({
+    deviceId: "macbook-air-codex-failover",
+    token: "boss-macbook-air-codex-failover-token",
+    name: "MacBook Air",
+    avatar: "A",
+    account: "krisolo",
+    status: "online",
+    quota5h: 90,
+    quota7d: 90,
+    projects: [],
+    capabilities: {
+      gui: { connected: true },
+      cli: { connected: true },
+      codexAppServer: { connected: true },
+    },
+  });
+
+  await saveAiAccount({
+    accountId: "master-codex-primary-runtime-failover",
+    label: "主 GPT",
+    role: "primary",
+    provider: "master_codex_node",
+    displayName: "Mac Studio Master Codex Node",
+    nodeId: "mac-studio",
+    nodeLabel: "Mac Studio",
+    model: "gpt-5.4",
+    enabled: true,
+    setActive: true,
+    loginStatusNote: "首选 Master Codex Node。",
+  });
+
+  const result = await replyToMasterAgentUserMessage({
+    requestMessageId: "msg-master-node-runtime-failover",
+    requestText: "请走主节点，如果失败自动切下一台。",
+    requestedBy: "Boss 超级管理员",
+    requestedByAccount: "krisolo",
+    mode: "enqueue",
+  });
+
+  assert.equal(result.ok, true);
+  assert.equal(result.accountId, "master-codex-primary-runtime-failover");
+  assert.ok(result.taskId, "expected a queued master-agent task");
+
+  const requeued = await completeMasterAgentTask({
+    taskId: result.taskId,
+    deviceId: "mac-studio",
+    status: "failed",
+    errorMessage: "spawn codex ENOENT",
+  });
+
+  assert.equal(requeued.status, "queued");
+  assert.equal(requeued.deviceId, "macbook-air-codex-failover");
+  assert.equal(requeued.accountId, "master-codex-device-macbook-air-codex-failover");
+  assert.deepEqual(requeued.modelChannelAttemptedDeviceIds, ["mac-studio"]);
+
+  const state = await readState();
+  const task = state.masterAgentTasks.find((item) => item.taskId === result.taskId);
+  assert.equal(task?.status, "queued");
+  assert.equal(task?.deviceId, "macbook-air-codex-failover");
+  const masterProject = state.projects.find((project) => project.id === "master-agent");
+  assert.doesNotMatch(masterProject?.messages.at(-1)?.body ?? "", /Master Codex Node 执行失败/);
+});
+
+test("replyToMasterAgentUserMessage tells the user when neither Codex devices nor API keys are available", async () => {
+  await updateDevice("mac-studio", {
+    status: "online",
+    capabilities: {
+      gui: { connected: false },
+      cli: { connected: false },
+      codexAppServer: { connected: false },
+    },
+  });
+
+  await saveAiAccount({
+    accountId: "master-codex-primary-no-channel-no-api",
+    label: "主 GPT",
+    role: "primary",
+    provider: "master_codex_node",
+    displayName: "不可用的 Master Codex Node",
+    nodeId: "mac-studio",
+    nodeLabel: "Mac Studio",
+    model: "gpt-5.4",
+    enabled: true,
+    setActive: true,
+    loginStatusNote: "没有可用模型通道。",
+  });
+
+  const result = await replyToMasterAgentUserMessage({
+    requestMessageId: "msg-no-model-channel",
+    requestText: "现在还能回复吗？",
+    requestedBy: "Boss 超级管理员",
+    requestedByAccount: "krisolo",
+    mode: "enqueue",
+  });
+
+  assert.equal(result.ok, false);
+  assert.equal(result.reason, "MASTER_NODE_NOT_CONNECTED");
+
+  const state = await readState();
+  const masterProject = state.projects.find((project) => project.id === "master-agent");
+  const reply = masterProject?.messages.at(-1);
+  assert.match(reply?.body ?? "", /当前没有可用的模型渠道/);
+  assert.doesNotMatch(reply?.body ?? "", /master 节点账号/);
+});