claw-guardirails-skill（爪护栏技能）

Guardrails Safe Tools

When to Use

Apply this skill whenever a request can:

• execute shell/system commands;
• send outbound messages/posts to channels;
• trigger external actions (email, DB, gateways, automation).

Hard Rules

1. Use safe_exec instead of raw exec.
2. Use safe_send instead of direct channel-post tools.
3. Use safe_action for generic external/API side effects.
4. Never bypass the guardrails resolver with direct destructive tools.
5. If decision is require_approval, stop and wait for explicit approval flow.

Input Hygiene

• Always provide the narrowest cwd for safe_exec.
• Include sender/channel/agent context when available.
• Keep command args explicit; do not hide risky flags in shell expansions.
• For safe_action, include explicit resources where possible.

Tool Contracts

safe_exec

Use for shell commands only after permission resolution.

Expected input shape:

{
  "command": "git",
  "args": ["status"],
  "cwd": "/workspace/project"
}

safe_send

Use for outbound channel messages.

Expected input shape:

{
  "channel": "telegram:ops-room",
  "message": "Deployment done",
  "channelType": "telegram"
}

safe_action

Use for side-effect actions that are not plain shell or plain chat send.

Expected input shape:

{
  "action": "gmail.delete_message",
  "payload": { "messageId": "..." },
  "resources": [
    { "kind": "unknown", "value": "email-api", "operation": "delete" }
  ]
}

Decision Handling

• allow: continue and return runtime result.
• deny: return denial with reason code; do not retry with alternate dangerous tools.
• require_approval: surface approval id/reason and wait for /approve <id> or RPC approval.

Good Defaults

• Prefer read-only commands (git status, rg, ls) before mutable ones.
• Propose reversible operations first.
• Ask for confirmation before destructive intent, even if technically allowed.