← 返回 Skills 市场
liuwujijay

Agi Term Helper

作者 liuwujijay · GitHub ↗ · v1.0.1
darwinlinuxwin32 ⚠ suspicious
836
总下载
0
收藏
2
当前安装
2
版本数
在 OpenClaw 中安装
/install agi-term-helper
功能描述
Fast, explicit terminal execution via OpenClaw exec (direct dispatch; you type the exact command).
使用说明 (SKILL.md)

/term — direct terminal execution (exec dispatch)

/term is a power-user shortcut: whatever you type after /term is forwarded as-is to OpenClaw’s exec tool.

This is intentionally “manual mode”:

  • You (the user) provide the exact shell command.
  • OpenClaw does not rewrite, expand, or “helpfully” change it.
  • It’s useful when you want quick, deterministic terminal actions without a planning loop.

How dispatch works (important)

OpenClaw supports command-dispatch: tool skills. When you run:

  • /term ls -la

the raw argument string (ls -la) is forwarded to the configured tool (exec) without extra parsing. In tool-dispatch mode, OpenClaw invokes the tool with params shaped like:

{ command: "\x3Craw args>", commandName: "\x3Cslash command>", skillName: "\x3Cskill name>" }

See the Skills docs section on command-dispatch, command-tool, and command-arg-mode. :contentReference[oaicite:10]{index=10}

When to use /term vs normal “agent runs”

Use /term when:

  • You already know the exact command you want.
  • You want a quick read-only check (files, git status, grep).
  • You are debugging OpenClaw itself (skills folder, logs, Peekaboo bridge status).

Prefer normal agent flow when:

  • You want the model to decide the best approach.
  • The task may need multiple steps, safety checks, or file edits.

Safety model (read this once, then follow it)

/term is equivalent to letting an assistant type into your terminal. Good defaults:

  1. Prefer read-only commands unless you mean to change state.
  2. Avoid secrets in command lines (tokens, API keys, cookies).
  3. Avoid remote execution one-liners:
    • no curl ... | sh
    • no “download and execute” pipelines
  4. If the command could delete or overwrite files, slow down and double-check paths.

Host + sandbox notes

Your actual execution environment depends on how you invoke exec in your setup (sandbox vs host). Also note: when a session is sandboxed, environment variables are not automatically inherited by the container; you must inject them via sandbox env settings or bake them into the image. :contentReference[oaicite:11]{index=11}

Practical examples

Quick inspection (safe, read-only)

  • /term pwd
  • /term ls -la
  • /term git status
  • /term rg -n "TODO|FIXME" .

Debug Peekaboo bridge discovery (macOS)

  • /term peekaboo bridge status --verbose

If the output shows “no such file or directory” for all candidates and “Selected: local (in-process)”, you likely have no bridge host running (see troubleshooting below).

One-liners for structured output

If you want JSON output for parsing:

  • /term python -c 'import json,platform; print(json.dumps({"py":platform.python_version()}))'

Troubleshooting

“command not found”

The tool runs in whatever PATH your OpenClaw runtime provides. If you rely on Homebrew, ensure the runtime sees /opt/homebrew/bin.

“permission denied” / macOS privacy prompts

Some tools (screen capture / UI automation) require Screen Recording / Accessibility. Those permissions are per-process context on macOS; using PeekabooBridge is often the reliable path for automation.

I need richer guidance and guardrails

Install/use the companion skill terminal-helper (model-invocable) which teaches safe patterns, confirmations, and runbooks.

安全使用建议
This skill is a thin passthrough: when you run /term <command>, that exact command is sent to your exec tool and executed in whatever environment OpenClaw provides. That behavior is intentional and internally consistent, but it also means you must treat it like typing directly into a terminal: never paste commands you don't understand, avoid putting secrets on command lines, and never run remote-download-and-execute one-liners (e.g., curl | sh). Because disable-model-invocation is true, the model won't run commands by itself — only you can trigger them — which reduces accidental execution risk. If you want safer, model-assisted checks and confirmations, consider installing a model-invocable companion (the SKILL.md suggests a 'terminal-helper').
功能分析
Type: OpenClaw Skill Name: agi-term-helper Version: 1.0.1 The `agi-term-helper` skill provides direct, unsanitized terminal execution via OpenClaw's `exec` tool, configured by `command-tool: exec` and `command-arg-mode: raw` in `SKILL.md`. While the documentation explicitly warns users about the risks of arbitrary command execution (e.g., avoiding `curl ... | sh`) and the skill is `disable-model-invocation: true` (preventing the AI agent from autonomously misusing it), this capability represents a significant security risk due to direct arbitrary command execution on the host system. This falls under 'risky capabilities without clear malicious intent' as per the classification guidelines.
能力评估
Purpose & Capability
Name/description claim a raw terminal passthrough and the skill requests no credentials, binaries, or installs — exactly what you'd expect for a thin /term -> exec dispatcher.
Instruction Scope
SKILL.md explicitly forwards the user's raw argument string to the configured exec tool and contains safety guidance (avoid secrets, avoid `curl | sh`, prefer read-only commands). This is coherent for a direct-exec helper, but by design it permits arbitrary shell commands (including destructive or exfiltratory commands) when invoked by a user. The skill itself does not instruct reading unrelated files or exfiltrating data.
Install Mechanism
No install spec and no code files are present (instruction-only). Nothing is written to disk and no external artifacts are fetched — lowest-risk install posture.
Credentials
The skill requires no environment variables or credentials. The doc notes that sandboxed runtimes may not inherit env and explains how to inject them; that explanation is informative and proportional.
Persistence & Privilege
always is false and disable-model-invocation is true (the model cannot invoke this autonomously). The skill does not request persistent system privileges or modify other skills' config.
如何使用
  1. 确保已安装 OpenClaw(本地或 Docker 部署)
  2. 在对话框中输入安装命令:/install agi-term-helper
  3. 安装完成后,直接呼叫该 Skill 的名称或使用 /agi-term-helper 触发
  4. 根据 Skill 的参数说明提供必要输入,即可获得结构化输出
版本历史
v1.0.1
- Added _meta.json file to the skill package. - No changes to skill behavior or documentation.
v1.0.0
Initial release of agi-term-helper v1.0.0: - Introduces the `/term` command for direct, explicit terminal execution via OpenClaw's `exec` tool. - Enables users to dispatch raw shell commands without model intervention or rewriting. - Designed for power users needing fast, deterministic command execution. - Documentation covers usage instructions, safety warnings, environment considerations, and troubleshooting tips. - Includes practical examples for common read-only and debugging scenarios.
元数据
Slug agi-term-helper
版本 1.0.1
许可证
累计安装 3
当前安装数 2
历史版本数 2
常见问题

Agi Term Helper 是什么?

Fast, explicit terminal execution via OpenClaw exec (direct dispatch; you type the exact command). 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 836 次。

如何安装 Agi Term Helper?

在 OpenClaw 或 Claude Code 对话框中运行命令「/install agi-term-helper」即可一键安装,无需额外配置。

Agi Term Helper 是免费的吗?

是的,Agi Term Helper 完全免费(开源免费),可自由下载、安装和使用。

Agi Term Helper 支持哪些平台?

Agi Term Helper 跨平台运行,可在任意部署了 OpenClaw / Claude Code 的环境中使用(darwin, linux, win32)。

谁开发了 Agi Term Helper?

由 liuwujijay(@liuwujijay)开发并维护,当前版本 v1.0.1。

💬 留言讨论