← 返回
未分类 中文

Sara: Safety Ordering Guard for AI Skills

Quiet logic guard for risky multi-skill workflows. Checks order before delete, send, booking, publishing, or other high-impact actions.
静默逻辑守卫,用于高风险多技能工作流。在删除、发送、预订、发布或其他高影响操作前检查顺序。
zc502 zc502 来源
未分类 clawhub v1.0.0 1 版本 100000 Key: 无需
★ 0
Stars
📥 355
下载
💾 0
安装
1
版本
#calendar#email#github#guardrails#latest#logic#orchestration#productivity#publishing#safety#workflow

概述

Sara

You have access to Sara, a quiet logic guard for multi-skill workflows.

Sara helps reduce risky ordering mistakes when a plan includes multiple skills or tool actions.

Use Sara especially when a plan includes:

  • destructive actions like delete, remove, clear, wipe, drop, truncate
  • external actions like send, publish, post, deploy, submit
  • booking or scheduling actions
  • repo-destructive actions
  • privacy, permission, or sensitive data access actions

What Sara checks

Sara v0.1 enforces a small set of hard safety rules:

  1. backup before delete
  2. check before operate
  3. permission before read
  4. preview before publish

Examples:

  • backup -> delete
  • check_calendar -> book_flight
  • review_draft -> send_email
  • request_access -> read_private_data
  • draft -> publish_post

When to use Sara

Run Sara before executing a risky multi-step plan when:

  • 2 or more skills/tools are involved, and
  • at least one step is high-risk, irreversible, externally visible, or destructive

For simple low-risk single-step actions, Sara is usually unnecessary.

Required behavior

Before executing a risky multi-step plan:

  1. Write down the proposed tool or skill order as a simple list.
  2. Run scripts/run_audit.py with that plan.
  3. Read the result carefully.
  4. If is_safe is false, do not continue with the original order.
  5. Show the user the risk and propose the suggested_order.
  6. If the action is destructive or irreversible, ask for confirmation before continuing.

Skipping Sara for risky multi-step plans increases the chance of contradictory actions, failed retries, or destructive ordering mistakes.

Example audit input

{"tools":["backup","delete"]}

Example audit output

{
  "is_safe": false,
  "risk_level": "critical",
  "warnings": [
    "Run 'backup' before 'delete'. Backups should happen before destructive actions."
  ],
  "suggested_order": ["backup", "delete"]
}

How to respond

  • If Sara says the plan is safe: continue.
  • If Sara says the plan is unsafe: pause, explain the risk, and suggest the safer order.
  • If the plan touches deletion, publishing, payment, scheduling, permissions, or sensitive data, prefer confirmation before continuing.

Response style

When Sara detects a risky order, explain it clearly and briefly.

Good example:

  • “Sara noticed a risky sequence. Suggested order: backup -> delete. Would you like me to continue in that order?”

Do not mention internal implementation details unless the user asks.

Important

Sara is a guard, not a replacement for user intent.

Use Sara to:

  • reduce risky ordering mistakes
  • reduce repeated retries
  • reduce contradictory multi-skill plans
  • make high-risk workflows safer and more consistent

版本历史

共 1 个版本

  • v1.0.0 当前
    2026-05-07 15:48 安全 安全

安全检测

腾讯云安全 (Keen)

安全,无风险
查看报告

腾讯云安全 (Sanbu)

安全,无风险
查看报告

🔗 相关推荐

ai-agent

Self-Improving + Proactive Agent

ivangdavila
自我反思+自我批评+自我学习+自组织记忆。智能体评估自身工作、发现错误并持续改进。
★ 1,444 📥 328,865
ai-agent

Agent Browser

rez0
用于 AI 代理的浏览器自动化 CLI。当用户需要与网站交互(包括浏览页面、填写表单、点击按钮、截图等)时使用。
★ 866 📥 347,130
ai-agent

Find Skills

root
帮助用户发现和安装智能体技能,当用户询问如「如何做X」、「找X的技能」、「有能做...的吗」等问题时
★ 1,519 📥 577,753