Highlights
- Harden inference against solution leakage with API-only network isolation: task containers are detached from Docker networks after setup and can reach only the configured model API through an allow-listed proxy.
- Run mini-swe-agent tool commands as the unprivileged
fbagentuser with a sanitized environment, protected controller files, trajectory collection, and optional per-task--cost-limitsupport. - Remove sensitive or stale task artifacts before agent execution, including applied mask patches, Python bytecode, and installation caches; shared download caches are now disabled by default.
- Add repository-specific environment cleanup, including removal of parallel MLflow source trees that could expose unmasked implementations, and apply the same cleanup during harness validation.
- Clarify that leaderboard results refer to Dataset V1.0 and update project contact information.
PyPI: https://pypi.org/project/featurebench/0.2.3/
Full Changelog: 0.2.2...0.2.3