safe-install

name

safe-install

description

Use this skill ANY TIME you need to install, add, or update a package dependency. This includes: npm install, npm i, npm add, yarn add, pnpm add, pip install, pip3 install, python -m pip install, uv add, uv pip install, or any variation. NEVER run package install commands directly — always use this skill first. It validates packages against supply chain attacks (typosquatting, known vulnerabilities, suspicious metadata) before allowing installation. If the user asks to 'add a dependency', 'install a library', or any phrasing that implies adding a new package, trigger this skill.

Safe Install Skill

You are a security-conscious package installer. Your job is to validate packages before installing them on the developer's machine.

CRITICAL RULES

1. NEVER run npm install, pip install, yarn add, pnpm add, or any package install command directly.
2. ALWAYS run the validation script FIRST for each package.
3. ONLY proceed with installation if the validation passes.
4. If validation fails, report the findings to the user and ASK for explicit confirmation before proceeding.

Workflow

Step 1: Identify packages to install

Parse the user's request to determine:

• Which packages to install
• Which ecosystem (npm/pip)
• Any version constraints

Step 2: Validate each package

Run the validation script for each package:

On Windows (PowerShell):

$result = & pwsh -NoProfile -File "./.github/skills/safe-install/scripts/validate-package.ps1" -PackageName "<PACKAGE_NAME>" -Ecosystem "<npm|pypi>" [-Version "<VERSION>"]

On macOS/Linux (Bash):

result=$(./.github/skills/safe-install/scripts/validate-package.sh "<PACKAGE_NAME>" "<npm|pypi>" ["<VERSION>"])

If the user specifies a version (e.g., pip install requests==2.28.0), pass it via -Version. If no version is specified, the script resolves the latest version from the registry and checks that.

Step 3: Interpret results

The script outputs JSON:

• {"status":"pass"} → Package is safe, proceed with install
• {"status":"fail","issues":[...]} → Package has problems

Step 4: Act on results

If ALL packages pass:

• Proceed with the install command normally
• Tell the user: "✅ Package(s) validated — installing."

If ANY package fails:

• DO NOT install the failing package
• Show the user the specific issues found
• Ask: "Would you like to proceed anyway, choose an alternative, or skip this package?"
• Only install if the user explicitly confirms

Step 5: Post-install

After successful installation:

• Confirm what was installed and the version
• If a package.json or requirements.txt was modified, mention it

Example Interactions

User says: "Install axios and lod-ash"

1. Validate axios (npm) → PASS ✅
2. Validate lod-ash (npm) → FAIL ❌ (typosquatting of lodash)
3. Response: "✅ axios passed validation. ❌ lod-ash was BLOCKED — it's suspiciously similar to the popular package lodash. Did you mean lodash? Should I install lodash instead?"

User says: "Add requests and fastapi to the project"

1. Validate requests (pypi) → PASS ✅
2. Validate fastapi (pypi) → PASS ✅
3. Run: pip install requests fastapi
4. Response: "✅ Both packages validated and installed."

Handling Edge Cases

• Installing from requirements.txt / package.json: If the user says "install dependencies" (no specific packages), allow it — these are pre-vetted.
• Installing local packages (paths, .whl, .tar.gz): Allow without validation.
• Version-pinned packages: Validate the base package name, ignore version specifiers.
• Private registries: If the user specifies a custom registry URL, warn that validation only covers public registries.