guard

Files (1)

• SKILL.md

SKILL.md (first lines)

Activate safety guard mode. Restricts editing scope and adds extra confirmation for destructive operations.

Scope: the task/scope the user described when invoking this skill (if none given, ask or infer from context)

## Modes
### Mode 1: Freeze (edit scope restriction)
When the task/scope the user described when invoking this skill (if none given, ask or infer from context) contains a path or file pattern:
- ONLY files matching the specified scope may be edited; all others are READ-ONLY this session.
- Before every edit, verify the target is within scope. If out of scope: warn and ask for explicit confirmation.
- Example: `$guard src/auth/` → only files under `src/auth/` may be modified.

### Mode 2: Careful (destructive operation warnings)
When the task/scope the user described when invoking this skill (if none given, ask or infer from context) is empty or contains "careful":
- Before any destructive shell command, print a warning and ask for confirmation:
  - `rm`, `git reset`, `git checkout .`, `git clean`
  - `DROP`, `TRUNCATE`, `DELETE FROM`
  - `docker rm`, `docker system prune`
  - Any command modifying files outside the project directory
- Before large-scale edits (5+ files), show a summary first.
- Before any git push, show what will be pushed.

### Mode 3: Full Guard (freeze + careful)
When the task/scope the user described when invoking this skill (if none given, ask or infer from context) contains both a path and "careful", apply both.

## Activation
Print:
```
Guard Mode Activated
- Scope: [restricted path or "all files"]
- Mode: [freeze / careful / full]
- To deactivate: tell me "guard off"
```

## During Session
Before every edit or shell command, check:
1. **Freeze check**: is the target within the allowed scope?
2. **Careful check**: is this destructive or large-scale?
3. If either fails → warn + ask confirmation before proceeding.

## Deactivation
When the user says "guard off" or "가드 해제":
```
Guard Mode Deactivated
- All restrictions removed
```

## Rules
- Guard mode is advisory — the user can always override with explicit confirmation.
- Don't be annoying — only warn on genuinely risky operations, not every edit.
- Always show what you're about to do when asking for confirmation.
- Track which files were modified during guard mode for the deactivation summary.

From Hbin77/codex-kit, MIT.

More from Hbin77/codex-kit