Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
34 lines (24 loc) · 1.36 KB

CONTRIBUTING.md

File metadata and controls

34 lines (24 loc) · 1.36 KB
Raw

Contributing

Bug reports, manifest questions and pull requests are welcome. For a sandbox bypass, use private reporting instead (see SECURITY.md).

Setup

git clone https://github.com/stanleys12/skilljail && cd skilljail
python3 -m venv .venv && . .venv/bin/activate
pip install -e '.[dev]'
skilljail doctor        # checks that the Seatbelt (macOS) or bwrap (Linux) backend works here

Tests

python -m pytest -q

Changes to the policy compiler, the proxy or the hooks should come with a test that fails without the change. The containment experiments are slower and run real jailed processes against a canary home directory, so they are not part of the unit suite:

skilljail eval --attack-suite eval/attack_suite --experiments attack,e4

The MalSkillBench experiments need the dataset checked out separately and pointed at with --malskillbench /path/to/MalSkillBench/Dataset/Skills; see docs/EVAL.md for how those numbers are produced.

Where things live

  • skilljail/manifest.py, policy.py: the manifest format and what it compiles to
  • skilljail/backends/: Seatbelt and bwrap profile generation
  • skilljail/proxy.py: the egress proxy and private-address checks
  • skilljail/hooks/claude_code.py: Claude Code hook handling
  • docs/SPEC.md: the manifest specification; update it when the format changes