Plugin Installation and Usage

Cerberus Plugin (Optional)

Cerberus provides secure command execution with policy-based sandboxing. It is included in the workspace but requires the eBPF toolchain (Linux only).

# Install with eBPF support (default, requires nightly Rust + eBPF toolchain)
cargo install --path crates/cerberus/cerberus-cli

# Install without eBPF if toolchain is unavailable
cargo install --path crates/cerberus/cerberus-cli --no-default-features -p cerberus-core

If cargo build --release fails due to Cerberus/eBPF, you can skip it:

cargo build --release --workspace --keep-going

Plugins

Pre-built hookers and skills are placed in <your_xiaoO>/plugins. They are not installed by default.

To install hookers, run:

cd <your_xiaoO>/plugins/hookers
./install.sh

You can also develop your own hookers and place them in <your_xiaoO>/plugins/hookers. See how-to-develop-a-plugin-hooker.md for details.

Skills

Built-in Skills

When you run cargo install --path apps/endside, builtin skills are automatically installed. They provide security policy enforcement and other built-in capabilities, and are loaded with highest priority by the runtime.

Installation locations (automatic fallback):

  • System level (preferred): /usr/lib/.xiaoo/skills/ - requires root privileges
  • User level (fallback): ~/.xiaoo/skills/ - used if system-level installation fails

Builtin skills (located in <xiaoO>/plugins/skills/):

  • xiaoo-guardian - Security policy enforcement
  • block-analyzer - Block analysis capabilities

Note: cargo build does NOT install skills. Only cargo install triggers skill installation.

Installation Behavior:

  • First attempts to install all builtin skills to system-level directory (requires root privileges)
  • If system-level installation fails (e.g., permission denied), automatically falls back to user-level directory
  • Without these skills, security features and other capabilities may be unavailable.

For system-wide installation (recommended for multi-user environments):

  • Run cargo install with root privileges: sudo cargo install --path apps/endside

Skill Directory Priority (Four Levels)

  1. Project level (highest): ./.xiaoo/skills/ - Project-specific skills
  2. Config level (medium): Directories specified in [skills].dirs - Team/user shared skills
  3. User level: ~/.xiaoo/skills/ - Personal skills available everywhere
  4. System level (lowest): /usr/lib/.xiaoo/skills/ - Built-in skills only

Custom Skills

Custom skills can be installed to user-level directory using the xiaoo --cli skill install command:

# Install from local directory (installs to ~/.xiaoo/skills by default)
xiaoo --cli skill install ./my-skill/

# Install from Git repository
xiaoo --cli skill install https://github.com/user/my-skill.git

Note: User-installed skills go to ~/.xiaoo/skills/ (user level), NOT /usr/lib/.xiaoo/skills/ (system level is reserved for built-in skills).

See skill_usage.md for detailed skill documentation.

Uninstalling Skills

# Remove a user-installed skill
xiaoo --cli skill remove <skill-name>

# Or manually remove from user level
rm -rf ~/.xiaoo/skills/<skill-name>

# Remove built-in guardian skill (requires root)
sudo rm -rf /usr/lib/.xiaoo/skills/xiaoo-guardian

To completely uninstall xiaoO and all associated skills:

# Uninstall the application
cargo uninstall xiaoo-endside

# Remove system-level skills (requires root)
sudo rm -rf /usr/lib/.xiaoo/skills

# Remove user-level skills
rm -rf ~/.xiaoo/skills

RPM 安装环境下的插件管理

通过 RPM 安装 xiaoO-hookers 后,所有插件安装到 /usr/lib/.xiaoo/hookers/ 目录下。插件可以通过以下两种方式启用和关闭。

方式一:使用软链接命令(推荐)

RPM 安装时会创建两个全局软链接,可直接在任意目录执行:

# 交互式安装/启用插件
xiaoo-hookers-install

# 卸载/禁用插件
xiaoo-hookers-uninstall

也支持非交互模式:

# 启用 audit_agent(不交互,直接写入配置)
xiaoo-hookers-install --non-interactive audit_agent

方式二:直接编辑 config.toml

手动编辑 ~/.config/xiaoo/config.toml[hooker] 段。

场景 1:启用所有插件

[hooker]
default = "All"
plugins = [
  "/usr/lib/.xiaoo/hookers/audit_agent/plugin.json",
]

default = "All" 表示所有已注册的插件默认启用,plugins 列表中的 plugin.json 会被加载注册。

场景 2:启用所有插件,但排除个别不需要的

[hooker]
default = "All"
plugins = [
  "/usr/lib/.xiaoo/hookers/audit_agent/plugin.json",
  "/usr/lib/.xiaoo/hookers/tool_post_secret_guard/plugin.json",
]
disabled = [
  "pluginguard_secret_like_output",
]

disabled 列表填写 plugin.json 中定义的 id 字段值。

RPM 场景下插件 install.sh 的变化

RPM 打包时,部分插件自身的 install.sh 行为会有所调整。以 audit_agent 为例:

  • 开发者场景(git clone + cargo build):执行 install.sh 时会创建 Python 虚拟环境(venv),并通过 pip 安装依赖
  • RPM 场景(通过 RPM 安装):audit_agent 的 Python 依赖已通过 RPM 包提供(python3-openaipython3-httpxpython3-pydanticpython3-tenacity 等),因此 install.sh 不再创建 venv 和 pip install,只负责生成 audit_settings.json 配置文件

这意味着在 RPM 场景下执行 xiaoo-hookers-install 不会因为缺少 pip 或 venv 而出错。

验证插件是否生效

# 查看 xiaoo 运行日志,确认插件已加载
xiaoo run -p "echo hello"

# 查看审计日志(audit_agent)
cat /usr/lib/.xiaoo/hookers/audit_agent/audit_policy_checker.log

注意事项

  • 路径使用绝对路径plugins 列表中的路径必须是 plugin.json 的完整绝对路径,RPM 安装后路径为 /usr/lib/.xiaoo/hookers/<plugin>/plugin.json
  • disabled 使用 hooker iddisabled 列表填写的是 plugin.json 中 id 字段的值(如 plugin_audit_tool_inputpluginguard_secret_like_output),不是文件路径
  • 修改配置后无需重启:xiaoo 每次启动时重新读取配置

Chat & Session Lifecycle Hooks

除 Tool 层 hook(*.Tool.*.pre/post/error)外,xiaoo 还提供 chat 层session 层 两个挂载点,让外部插件能在用户输入、system prompt 组装、会话状态流转等阶段介入。它们与 Tool hook 共用同一套 plugin.json + sh -c 子进程 + stdin/stdout JSON 协议,可注册在同一个 ~/.xiaoo/hookers/<name>/plugin.json 数组里。

Hook points overview

Hook point 触发时机 插件可返回结果
*.Chat.command.before 用户输入命中 ~/.xiaoo/commands/<name>.md 斜杠命令、模板展开为 body 之后、body 提交为 user turn 之前 Allow / Transform { body } / Deny { reason }
*.Chat.message.received user 消息构造完成、写入消息历史之前 Accept / Transform { message }
*.Chat.system.transform PromptBuilder 组装完 system: Vec<String> 分段、合并成单条 system 消息之前 Allow / Transform { system }
*.Session.lifecycle.state 一次非错误 root turn 结束(Complete/MaxTurnsReached/BudgetExhausted/Cancelled 四种 Ok 结局)、会话回到 idle 时(fire-and-forget,不阻塞 turn 返回) Ack(事件型,无可变输出)

前三个 chat hook 是「可变 hook」——插件可以改写或拒绝输入;第四个 session hook 是「事件型观察者」——只能确认收到事件,没有 transform/deny 路径。

与 Tool hook 的差异

  • 执行模型相同:都是 sh -c <command>,stdin 写入一次 JSON payload、stdout 读取一次 JSON 结果,非零退出视为失败。
  • payload.stage 不同:chat/session hook 的 stage 字符串是 command_before / chat_message / system_transform / session_state(不是 pre/post/error),脚本据此分发。
  • 调度差异:chat hook 在 agent loop 同步执行,单个插件报错只 tracing::warn! 不中断整轮,只有 command.beforeDeny 会短路;session state hook 在 gateway 后台 tokio::spawn 执行(fire-and-forget),错误走 tracing::warn!,绝不影响主流程或 run_turn 返回值。
  • 交互机制:三个 chat hook 还支持 action: "ask_user"——插件可发起 Confirm / TextInput / Choice 交互,用户回答后 xiaoo 会带着 interaction 字段再次调用同一命令,直到插件返回 final。session state hook 不支持该机制。

Minimal plugin.json

~/.xiaoo/hookers/js-test/plugin.json 一次性注册四个 hooker,全指向同一个脚本:

[
  { "id": "js_test_system_transform", "hook_point": "*.Chat.system.transform",    "command": "node ~/.xiaoo/hookers/js-test/xiaoo-hook-test.js" },
  { "id": "js_test_chat_message",     "hook_point": "*.Chat.message.received",   "command": "node ~/.xiaoo/hookers/js-test/xiaoo-hook-test.js" },
  { "id": "js_test_command_before",   "hook_point": "*.Chat.command.before",     "command": "node ~/.xiaoo/hookers/js-test/xiaoo-hook-test.js" },
  { "id": "js_test_session_state",    "hook_point": "*.Session.lifecycle.state", "command": "node ~/.xiaoo/hookers/js-test/xiaoo-hook-test.js" }
]

stage 取值表

payload.stage 是 xiaoo 在构造 payload 时硬编码写入的判别字符串,不等于 hook_point 的最后一段:

hook_point(plugin.json 里写的) payload 里的 stage
*.Chat.system.transform system_transform
*.Chat.message.received chat_message
*.Chat.command.before command_before
*.Session.lifecycle.state session_state

session_id 获取

插件需要当前会话 id 时,直接读 payload.session_id 即可,无需区分 local / remote 模式

  • local 模式:TUI 进程内跑 agent loop,hooker 子进程由 TUI 拉起,payload.session_id 就是 TUI 的 state.session_id(UUID)。
  • remote 模式:TUI 把自己的 session_id 通过 RuntimeTurnRequest 发给 daemon,daemon 原样使用;hooker 子进程由 daemon 拉起,payload.session_id 与 local 模式下是同一个值

因此插件不必判断"我现在是被 TUI 还是 daemon 拉起的",payload.session_id 永远是「当前会话」的正确 id。

各 hook 点 payload.session_id 可用性:

hook_point payload.session_id 类型
*.Chat.command.before String
*.Chat.message.received String
*.Chat.system.transform String | null
*.Session.lifecycle.state String
*.Tool.*.pre String
*.Tool.*.post / error payload 不带 session_id
*.Llm.*.pre / post / error payload 不带 session_id

system_transform 的 session_id 理论上可能为 nullChatSystemTransformInput.session_idOption<String>),脚本里建议 payload.session_id || "(unknown)" 兜底。tool post/error 与 LLM hook 的 payload 目前不带 session_id;如需在这些钩子里拿会话 id,可考虑改用 *.Tool.*.pre 或 chat 层钩子记录映射。

Minimal demo script

最简 demo(.js):每个 hook 都演示如何获取 session id 并做一次用户可感知的改写——[sid=<session_id>] 标记会出现在用户消息、system 指令和 /tmp/xiaoo-demo.log 日志里,便于核对每个钩子各自拿到的会话 id(command_before 额外拼 payload.command/payload.argumentschat_message 额外拼 payload.prior_message_count 判定首条消息):

#!/usr/bin/env node
// xiaoo hooker demo (js-test): 在每个 hook 点演示如何获取当前 session id。
// 直接读 payload.session_id 即可,无需区分 local/remote 模式(两种模式下
// 是同一个值——TUI 把 session_id 发给 daemon,daemon 原样使用)。
// stdout 只写结果 JSON;日志走 stderr 或文件。

const fs = require("fs");
const payload = JSON.parse(fs.readFileSync(0, "utf8") || "{}");

// 统一取 session_id:system_transform 的该字段可能为 null,做兜底。
const SID = payload.session_id || "(unknown)";

// 每个 hook 触发时都往日志写一行,便于核对四个钩子各自拿到的 sid。
fs.appendFileSync("/tmp/xiaoo-demo.log",
  `[${new Date().toISOString()}] stage=${payload.stage} sid=${SID}\n`);

let result;
switch (payload.stage) {
  case "command_before": {
    // payload.session_id = 当前会话 id
    // payload.command    = 斜杠命令名(如 "hook-test"),仅斜杠命令入口进此 stage
    // payload.arguments  = 原始参数串(如 "aaa",无参为空串)
    // payload.body       = 模板展开后的 body
    const cmd = payload.command || "";
    const args = payload.arguments || "";
    const tag = cmd ? `[HOOK:command_before:${cmd}${args ? ` ${args}` : ""}]` : "[HOOK:command_before]";
    result = { result: "transform", body: `${tag} [sid=${SID}] ${payload.body || ""}` };
    break;
  }
  case "chat_message": {
    // payload.session_id = 当前会话 id
    // payload.prior_message_count = 当前用户消息落库前的历史消息数:
    //   全新 session 首条 = 0;首轮完成后再发 = 2;只回放了 user 的恢复 = 1。
    //   用 <= 1 判定「会话首条有效输入」(兼容 retry/中断恢复)。
    const count = payload.prior_message_count ?? 0;
    const tag = count <= 1
      ? `[HOOK:chat_message:first#${count}]`
      : `[HOOK:chat_message:follow-up#${count}]`;
    const blocks = (payload.message?.blocks || []).map((b) =>
      b.type === "text" ? { ...b, text: `${tag} [sid=${SID}] ${b.text}` } : b
    );
    result = { result: "transform", message: { ...payload.message, blocks } };
    break;
  }
  case "system_transform": {
    // payload.session_id = 当前会话 id(可能为 null,已在顶部兜底)
    // 往 system 数组追加一条指令 → LLM 回复会以 [HOOK:system_transform] 开头
    const sys = Array.isArray(payload.system) ? payload.system : [];
    result = { result: "transform", system: [...sys, `回复必须以 [HOOK:system_transform] 开头(sid=${SID})`] };
    break;
  }
  case "session_state": {
    // payload.session_id = 当前会话 id(事件型、只读)
    // payload.state      = "idle"(一轮非错误结束)
    // payload.outcome    = complete / max_turns_reached / budget_exhausted / cancelled
    if (payload.state === "idle" && payload.outcome === "max_turns_reached") {
      const newSession = `cont-${Date.now()}`;
      result = {
        result: "ack",
        actions: [
          { kind: "create_session", session_id: newSession },
          { kind: "switch_session", session_id: newSession }
        ]
      };
    } else {
      result = { result: "ack" };
    }
    break;
  }
  default:
    result = { result: "allow" };
}

// 当某个 case 被注释掉、result 仍为 undefined 时,按 stage 回退到对应
// 的 no-op 结果,避免 JSON.stringify(undefined) → process.stdout.write(undefined)
// 抛 ERR_INVALID_ARG_TYPE 让整个 hooker 退出状态 1、被宿主当成失败跳过。
if (result === undefined) {
  const noop = {
    command_before: { result: "allow" },
    chat_message: { result: "accept" },
    system_transform: { result: "allow" },
    session_state: { result: "ack" },
  };
  result = noop[payload.stage] ?? { result: "allow" };
}
process.stdout.write(JSON.stringify(result));

运行后核对 session id:

# 四个钩子每次触发都会写一行,可直接看到各钩子拿到的 sid
cat /tmp/xiaoo-demo.log
# [2026-...Z] stage=command_before sid=550e8400-e29b-...
# [2026-...Z] stage=chat_message sid=550e8400-e29b-...
# [2026-...Z] stage=system_transform sid=550e8400-e29b-...
# [2026-...Z] stage=session_state sid=550e8400-e29b-...

local 与 remote 模式下 sid 值一致;remote 模式下日志写在 daemon 所在机器(hooker 由 daemon 进程拉起),不是 TUI 机器。

完整的 payload 字段、各 hook 合法输出 JSON 形状、ask_user 交互协议细节,请参考 plugins/hookers/how-to-develop-a-plugin-hooker.md 中的 Chat hook 与 Session lifecycle state hook 章节。

Plugin-requested session actions

*.Session.lifecycle.state hook 的响应里除了 result: "ack",还可以携带一个 actions 数组,让插件请求宿主在 turn 结束之后执行副作用——目前支持两种:

kind 必填字段 daemon 侧行为 TUI 侧行为
create_session session_id 调用 open_session(幂等恢复)打开/恢复该会话 切换焦点到该会话并恢复历史 transcript
switch_session session_id 调用 open_session(幂等恢复)确保目标会话存在 切换焦点到该会话并恢复历史 transcript

响应 JSON 形状:

{
  "result": "ack",
  "actions": [
    { "kind": "create_session", "session_id": "debug-1" },
    { "kind": "switch_session", "session_id": "debug-1" }
  ]
}

要点:

  • 只有 *.Session.lifecycle.state hook 的 adaptor 会解析 actions;chat/tool hook 的响应里写 actions 会被忽略。
  • session state hook 从 fire-and-forget 改为 awaited:dispatcher 会等所有 hooker 跑完、收集 actions 后才把 Done 事件发给 TUI,所以慢的 hooker 会延迟 Done——actions-emitting hooker 务必保持快。
  • daemon 侧先 open_session,失败的 action 被 tracing::warn! 记录并过滤,不会转发到 TUI;TUI 侧再异步切换焦点并恢复 transcript。
  • 整条链路 best-effort:任何一层失败都不会向 hook 调用者回传错误。
  • kindsnake_case,未知 kind 或缺字段的条目会被静默跳过——单个坏条目不会污染数组其它条目。
  • 宿主强制 max action depth = 3hook→action→hook 递归:超过上限的批次会被截断(保留前 3 条,多余的条目被 tracing::warn! 记录后丢弃)。插件作者请按此硬上限设计。

完整字段定义、daemon/TUI 两端执行流程、最小示例与排错清单,请参考 plugins/hookers/how-to-develop-a-plugin-hooker.md 第 16 章。