Pre-Action Guard

SKILL-018: Pre-Action Guard

Overview

Enforces a "Measure Twice, Cut Once" philosophy. This skill intercepts high-stakes actions (file writes, command execution) and validates them against safety rules and the current active plan to prevent accidental data loss or system corruption.

Trigger Phrases

• check safety
• verify action
• is this safe
• guard check

Inputs

| Parameter | Type | Required | Default | Description | |-----------|------|----------|---------|-------------| | --Action | string | Yes | - | The tool/verb being performed (e.g. write_to_file, rm) | | --Target | string | Yes | - | The file path or command argument | | --Plan | string | No | $null | The current plan context for alignment check |

Outputs

1. Assessment Result (JSON)

{
  "allowed": false,
  "risk_level": "CRITICAL",
  "reason": "Root directory deletion detected.",
  "warnings": ["Target matches blocked pattern '/*'."]
}

2. Risk Levels

• NONE: Read-only operations.
• LOW: Safe writes (files in temp, new files).
• MEDIUM: Modifying existing non-critical files.
• HIGH: Modifying config/env files.
• CRITICAL: Destructive delete/format commands.

Preconditions

1. Valid tool/action inputs.
2. PowerShell 5.1+ or Core 7+.

Safety/QA Checks

1. Auto-Block: Root deletions (/, C:\) are automatically blocked.
2. Context Awareness: Checks .env and config keywords for higher sensitivity.
3. Command-Only Discipline:

• Rejects output containing shell prompts (PS C:\, >).
• Rejects lines that look like terminal output (e.g., Everything up-to-date, At line:, CategoryInfo:).
• Rejects mixed snippets containing both commands and output.
• Enforcement: All runnable commands must be isolated in a clean, validated block.

Stop Conditions

| Condition | Action | |-----------|--------| | Missing Action/Target | Return error | | CRITICAL risk | Return allowed: false | | Mixed Output Detected | Return allowed: false |

Implementation

See scripts/guard_check.ps1.

Integration with Other Skills

All agents must:

1. Call this skill BEFORE running write_to_file on existing files.
2. Call this skill BEFORE running any run_command containing rm, del, drop.
3. If allowed is false, STOP and ask user for confirmation.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Sounder25
• Source: Sounder25/Foundational-Agent-Skills
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.