Back to skills
SKILL.md
Zero Cat Bash Prevention
ASecurityUse bounded native file readers when available to inspect source with line numbers and relevant context.
- 2 stars
- 0 votes
- 0 copies
- 0 views
- Added October 22, 2026
Works with
Security analysis
100/100npx -y skills add alivirgo/Major-AI-Skills --skill zero-cat-bash-prevention --agent claude-codeAre you the author of Zero Cat Bash Prevention?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/alivirgo-zero-cat-bash-prevention)---
name: zero-cat-bash-prevention
description: "Use bounded native file readers when available to inspect source with line numbers and relevant context."
category: efficiency
risk: safe
source: self
source_type: self
date_added: "2026-08-26"
tags: ["zero-cat", "view-file", "native-tools", "line-numbering", "token-optimization", "agent-architecture"]
tools: ["claude", "cursor", "gemini", "codex", "lmstudio"]
---
# Zero-Cat Native Inspection Protocol (view_file over Shell cat)
## Overview
When inspecting a file on disk (*e.g., `src/services/billing.ts`*), naive agents frequently spawn terminal subprocesses to execute Unix or Windows print commands (*`run_command("cat src/services/billing.ts")`* or *`run_command("type file.ts")`*).
Using shell `cat` introduces four major operational failures:
1. **Zero Line-Number Indexing**: Shell `cat` outputs raw unnumbered text. The agent cannot accurately determine `StartLine` or `EndLine` for downstream `replace_file_content` edits, causing hallucinated line offsets.
2. **Zero Slice Bounding**: `cat` dumps the entire 2,000-line file into context, burning **15,000+ tokens** when the agent only needed to check a 10-line function.
3. **Subprocess Process Latency**: Initializing a shell subprocess takes 400ms+, whereas native IDE filesystem APIs execute in **sub-millisecond memory reads**.
4. **Cross-Platform Incompatibility**: `cat` fails or produces errors on Windows `cmd.exe` environments.
The **Zero-Cat Native Inspection Protocol** mandates using the native **`view_file`** tool, returning **1-indexed line numbers (`42: code`)** with precise `[StartLine, EndLine]` slice controls.
---
## Shell Subprocess `cat` vs. Native `view_file` Tool
```
┌─────────────────────────────────────────────────────────────┐
│ File Inspection Mechanics │
│ │
│ Shell Subprocess `cat file.ts` (600 Lines / 4,800 Tokens): │
│ import { db } from './db'; │
│ export function query() { ... } │
│ // [595 unindexed lines dumped into context] │
│ ↳ 4,800 tokens billed, 0 line numbers for editing! │
│ │
│ Native `view_file` Tool (30 Lines / 280 Tokens - 94.1% Cut):│
│ 40: export async function chargeCustomer(id: string) { │
│ 41: const customer = await getCustomer(id); │
│ 42: if (!customer) throw new NotFoundError(); │
│ ↳ 280 clean tokens, exact line numbers ready for patching │
└─────────────────────────────────────────────────────────────┘
```
---
## The 4 Built-In Superpowers of `view_file`
```
┌───────────────────────────────────────────────────────────────────────────┐
│ 1. 1-INDEXED LINE NUMBERING: Every line prefixed with `<line_no>: ` │
│ 2. SURGICAL SLICE WINDOWING: `StartLine` and `EndLine` parameters │
│ 3. SUB-MILLISECOND VFS READS: Reads from IDE memory buffer without PTY │
│ 4. CROSS-PLATFORM DETERMINISM: 100% Identical on Windows, macOS, Linux │
└───────────────────────────────────────────────────────────────────────────┘
```
---
## Tool Invocation Standard
When inspecting code in Antigravity IDE / Claude Code:
```json
{
"AbsolutePath": "c:/Users/ASUS/Documents/Newfolder/Antigravity/Major AI Skills/src/services/billing.ts",
"StartLine": 40,
"EndLine": 75,
"toolAction": "Inspecting customer billing handler slice",
"toolSummary": "Surgical File View"
}
```
---
## Benchmark Comparison
Inspecting and editing 50 target functions in full-stack repositories:
| Inspection Tool | Ingested Tokens / Inspection | Line Number Indexing | Subsequent Edit Success |
| :--- | :--- | :--- | :--- |
| **Shell `cat` / `Get-Content`** | 4,200 tokens | ❌ None (Unnumbered) | 52% (Frequent line offset errors) |
| **Native `view_file` Tool** | **310 tokens** | **✅ Exact 1-Indexed** | **98% (First-pass patch success)** |
---
## Agent Operational Directive
> **MANDATORY**: Agents must NEVER run `cat`, `type`, `head`, `tail`, or `Get-Content` via shell tools (`run_command`). Always call the dedicated native `view_file` tool with specific `StartLine` and `EndLine` arguments.
Attribution
Comments
Loading comments…
Newsletter
Get the weekly list of safe new skills, and the ones we pulled.