antigravity-vscode: Skill for Claude Code

Code Security Checker

A comprehensive security audit loop: detect scope → review → fix → test → (optionally) embed in test suite. Maximum 30 fix iterations before escalating to manual intervention.

---

Step 1 — Determine scope

Run:

git diff HEAD

If the diff is non-empty: scope = changed files only. Extract the list:

git diff HEAD --name-only

Announce: "Git changes detected. Reviewing changed files only: <list>."

If the diff is empty: ask the user:

No git changes detected. Would you like to run a full-codebase security review? (yes / no)

• no → stop here, nothing to do.
• yes → scope = entire repository. Announce: "Running full-codebase security review."

---

Step 2 — Security review

Analyse only the in-scope files (changed files, or all source files for a full review). Apply the methodology below.

What to look for

Injection

• Command injection: user-controlled data passed to spawn, exec, shell strings, template literals that build shell commands.
• Path traversal: user input joined to a file path without normalisation/sandboxing (path.join with untrusted parts, readFile(userInput)).
• XSS: innerHTML, outerHTML, document.write, dangerouslySetInnerHTML set from external data without escaping.
• SQL / NoSQL injection: string interpolation into queries.

Authentication & authorisation

• Auth bypass: missing checks, hardcoded credentials, predictable tokens.
• Privilege escalation: logic that lets a lower-privilege actor reach a higher-privilege path.

Secrets & crypto

• Hardcoded API keys, passwords, tokens, or private keys in source.
• Weak algorithms (MD5/SHA1 for passwords, Math.random() for security tokens).

Data exposure

• Sensitive fields (passwords, tokens, PII) written to logs or error messages.
• Overly broad API responses that return internal fields.

Deserialization

• eval, new Function, unsafe JSON.parse of untrusted payloads, pickle.loads, yaml.load (unsafe).