Add Python script static analysis handler

[ldayton]

Summary

Adds a handler that auto-approves safe Python scripts by statically analyzing their AST. Uses a strict whitelist approach.

Threat Model

This is NOT a sandbox. We're protecting against AI coding assistants (Claude Code) accidentally running dangerous scripts, not against adversarial or malicious code.

Key insight: An AI won't deliberately craft tokenize.open() bypasses or ().__class__.__bases__[0].__subclasses__() exploits. It will accidentally reach for os.system(), subprocess.run(), or open("file.txt", "w").

What gets auto-approved

python --version              # Version/help flags
python -m json.tool           # Safe stdlib modules
python safe_script.py         # Scripts that pass analysis

Safe modules (pure computation, no I/O):
json, re, math, collections, itertools, functools, hashlib, datetime, dataclasses, typing, ast, zlib, etc.

What requires confirmation

python                        # Interactive mode
python -c 'code'              # Inline code
python -m http.server         # Arbitrary modules
python script_with_os.py      # Dangerous imports

Blocked modules (likely AI mistakes):

Module	Risk
os, subprocess, shutil	System execution
pathlib, io, open()	File I/O
socket, http, urllib	Network
gzip, bz2, lzma	Hidden file I/O via .open()
codecs	Hidden file I/O via .open()
inspect	File reading via getsource()

Blocked builtins: eval, exec, open, __import__, globals, getattr

Conservative defaults

• Unknown imports → ask (third-party packages like pandas, requests)
• Large files (>100KB) → ask
• Syntax errors → ask
• Non-.py files → ask

Test plan

• 67 tests covering safe scripts, dangerous scripts, bypass attempts
• just check passes on Python 3.11-3.14