权限模式与模型状态变更协议

Agent 权限模式 (agent_full_access / agent_activity) 和模型状态变更 (agent_model_state_update) 协议分析。

---

权限模式体系

模式等级

flowchart LR
    subgraph PermissionLevels["权限等级"]
        RO["read-only<br/>只读模式"]
        AE["auto-edit<br/>自动编辑"]
        AG["agent<br/>Agent 模式"]
        FA["full-access<br/>全权限模式"]
        YO["yolo<br/>绕过所有权限"]
    end

    RO -->|升级| AE
    AE -->|升级| AG
    AG -->|升级| FA
    FA -->|升级| YO

模式映射

// source: host/index.js — BA (normalizeModeToPermissionLevel)
function normalizeModeToPermissionLevel(mode, platform) {
    switch(mode) {
        // 只读模式
        case "read-only":
        case "read_only":
            return "plan";

        // 自动编辑
        case "auto_edit":
        case "auto-edit":
        case "autoEdit":
        case "accept_edits":
            return "autoEdit";

        // Agent 模式
        case "agent":
            return platform === "codex" ? "default" : undefined;

        // 全权限模式
        case "agent-full-access":
        case "agent_full_access":
        case "full-access":
        case "full_access":
            return "bypassPermissions";

        // 完全绕过
        case "yolo":
        case "bypassPermissions":
        case "dontAsk":
            return "yolo";

        // 预设方案
        case "plan":   return "plan";
        case "build":  return "plan";
        case "auto":   return "auto";
    }
}

模式与事件关系

模式	内部值	说明
read-only	plan	只能读文件，不能修改
auto-edit	autoEdit	自动编辑已存在的文件
agent	default / 空	标准 Agent 模式
full-access	bypassPermissions	完整权限，会弹权限请求
yolo	yolo	绕过所有权限确认，完全自动

---

agent_full_access 事件

当用户切换到 full-access 模式时，Agent 发出该事件：

{
    type: "agent_full_access",
    // 无额外 payload，纯信号事件
}

---

权限请求协议

sequenceDiagram
    participant Agent as ACP Agent
    participant Host as Host Process
    participant User as 用户

    Agent->>Agent: 需要执行敏感操作
    Agent->>Host: permission_request
    Note over Agent,Host: {requestId, description, kind, title, options, raw}

    Host->>User: 显示权限请求弹窗
    User->>Host: 选择允许/拒绝
    Host->>Agent: permission_response
    Note over Host,Agent: {requestId, optionId, response: {decision: "allow"|"deny"}}

    alt decision === "allow"
        Agent->>Agent: 执行操作
    else decision === "deny"
        Agent->>Host: tool_call_update (status: "failed")
    end

permission_request payload

{
    type: "permission_request",
    taskId: "task_xxx",
    traceId: "...",
    requestId: "req_xxx",
    description: "执行命令: rm -rf /",
    kind: "bash",            // 工具类型: bash / file / command
    title: "Bash",           // 显示标题
    options: ["allow", "deny", "allowOnce"],  // 可用选项
    raw: { ... }             // 原始请求数据
}

permission_response payload

{
    type: "permission_response",
    taskId: "task_xxx",
    traceId: "...",
    requestId: "req_xxx",
    optionId: "allow",
    response: {
        decision: "allow"    // "allow" | "deny"
    }
}

权限响应 → tool_call_update 联动

当用户拒绝时，Host 不仅回复 permission_response，还会发送 tool_call_update：

// 批准时
[{ type: "permission_response", decision: "allow" }]

// 拒绝时
[
    { type: "permission_response", decision: "deny" },
    { type: "tool_call_update", status: "failed", error: "..." }
]

---

agent_model_state_update 事件

当 Agent 的模型或推理级别发生变化时发出：

// source: host/index.js — agent_model_state_update dispatch
{
    type: "agent_model_state_update",     // 注意：可能前缀为 glm_
    taskId: "task_xxx",
    traceId: "...",
    version: 1,
    sessionId: "task_xxx",
    reason: "thought_level_changed"       // 变更原因
        | "model_changed"                 // 模型切换
        | "session_initialized",          // 会话初始化
    model: {
        currentValue: "glm-5.1"           // 当前模型
    },
    thoughtLevel: {
        enabled: true,                    // 是否启用推理
        budgetTokens: 2048                // 推理 token 预算
    }
}

reason 枚举

值	说明
thought_level_changed	推理层级变更（开关/预算调整）
model_changed	模型切换（如从 GLM-5.1 切到 GLM-5-Turbo）
session_initialized	会话首次初始化

---

agent_activity 事件

Agent 的结构化活动通知（含 thought 字段）：

// source: host/index.js — Q4 (parseAgentActivityResultContent)
{
    kind: "agent_activity",
    content: "...",           // 活动描述文本
    thought: "..."            // 可选，Agent 的思考过程
}

解析逻辑：

function Q4(agentType, rawContent) {
    // 1. 解析 JSON
    let parsed = parseJSON(rawContent);
    if (!parsed) return null;

    // 2. 验证 agentId / agentType
    if (agentType !== "Agent" && !parsed.agentId && !parsed.agentType)
        return null;

    // 3. 提取 content
    let content = extractText(parsed.content);
    if (!content) return null;

    // 4. 提取 thought
    let thought = parsed.thought;

    return {
        kind: "agent_activity",
        content: content,
        ...(thought ? { thought } : {})
    };
}

---

事件关系总览

graph TB
    subgraph Permission["权限系统"]
        PERM_REQ["permission_request<br/>请求权限"]
        PERM_RES["permission_response<br/>权限响应"]
        FULL_ACC["agent_full_access<br/>全权限模式"]
        ACTIVITY["agent_activity<br/>活动通知"]
    end

    subgraph State["状态更新"]
        STATE_UP["agent_model_state_update<br/>reason: thought_level_changed"]
        STATE_MODEL["agent_model_state_update<br/>reason: model_changed"]
        STATE_INIT["agent_model_state_update<br/>reason: session_initialized"]
    end

    subgraph Tool["工具执行"]
        TCALL["tool_call / tool_call_update"]
    end

    PERM_RES -->|"deny"| TCALL
    PERM_RES -->|"allow"| Tool
    STATE_UP -->|"推理调整"| Tool
    STATE_MODEL -->|"模型切换"| Tool

---

API 端点

支付相关（本轮发现的）

端点	方法	说明
/pay/create-sign	POST	支付宝/微信支付签名
/pay/check	GET	查询支付状态
/stripe/query	GET	查询 Stripe 已绑卡
/stripe/bind	POST	Stripe 绑卡
/stripe/pay	POST	Stripe 支付/订阅
/paypal/isSupport	GET	PayPal 可用性
/paypal/setupToken	POST	PayPal 授权 token
/paypal/subscribe	POST	PayPal 订阅

---

关键代码索引

函数	位置	说明
BA()	host/index.js	权限等级归一化
Q4()	host/index.js	agent_activity 解析
permissionRequestToStreamEvent()	host/index.js	n6, r6 函数
permission_response dispatch	host/index.js	权限响应处理
glm_agent_model_state_update	host/index.js	模型状态变更事件
normalizeModeToPermissionLevel()	host/index.js	模式→等级映射