WFC:SAFEGUARD - Write-Time Security Pattern Detection

Intercepts dangerous code patterns before tool calls complete. This is a
lint-level regex detector — it does NOT perform AST analysis and will produce
false positives on comments, strings, and documentation containing blocked
patterns.

Installation

This skill appends a PreToolUse hook entry to .claude/settings.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Write|Edit|Bash",
        "hooks": [
          {
            "type": "command",
            "command": "python3 wfc/scripts/hooks/safeguard.py"
          }
        ]
      }
    ]
  }
}

Merge behavior: If settings.json exists, the hook is appended to any
existing PreToolUse hooks. Deduplication is NOT performed — running this skill
multiple times will add duplicate entries.

File requirements: .claude/settings.json must be standard JSON.
JSONC (with comments) is NOT supported and will cause corruption or parse
failures.

Prerequisites Verification

Before installation, verify:

1. python3 command exists in PATH
2. wfc/scripts/hooks/safeguard.py exists and is executable
3. .claude/settings.json is writable (or creatable)

If any prerequisite fails, abort installation and report the specific failure.

What It Detects

Blocked Patterns (exit code 2 — tool call prevented)

Pattern	Languages	Notes
eval(	JS/TS, Python	Matches any use including in comments/strings
new Function(	JS/TS	Dynamic code execution
os.system(	Python	Command execution
subprocess with shell=True	Python	Shell injection vector
rm -rf /, rm -rf /*, rm -rf ~/	Bash	System destruction
${{ github.event.* }} in run: blocks	YAML	Actions expression injection

Warned Patterns (message to stderr — tool call proceeds)

Pattern	Languages	Risk
.innerHTML	JS/TS	XSS
dangerouslySetInnerHTML	React	XSS
pickle.load(	Python	Deserialization
child_process.exec(	Node.js	Command injection

Note: Hardcoded secret detection is NOT included in this version due to high
false-positive rates. Use wfc-security for secret scanning.

Known Limitations

1. Context blindness: Will block # TODO: remove eval() as if it were code
2. No AST parsing: Cannot detect indirect/obfuscated patterns
3. No deduplication: Re-running adds duplicate hooks
4. No uninstall: Removing hooks requires manual settings.json edit
5. JSON only: Cannot handle JSONC configuration files

Troubleshooting

Problem	Solution
Hook fails silently	Verify python3 in PATH; check script exists
False positives on docs	Expected behavior — disable hook to write docs
Duplicate hooks	Manually edit settings.json to remove extras
Can't modify settings	Check file permissions; ensure not read-only
JSON parse errors	Remove comments from settings.json