# OpenBench SWARM Agent Guidelines

- Architecture: MVP focusing on a privacy-preserving KPI share/aggregation pipeline with offline-first storage.
- Tech Stack: Python 3.8+, standard library, no heavy dependencies for MVP; packaging via setuptools; tests with pytest.
- Testing: `pytest` for unit tests; `python3 -m build` to verify packaging metadata and directory structure.
- Running tests: `bash test.sh` in the repo root.
- Contribution Rules: one feature per patch; keep changes minimal; avoid touching unrelated areas.
- Data Model: KPIRecord with revenue, COGS, inventory_turns, lead_time, CAC, LTV; anonymous sharing via anon_id.
- Privacy: aggregate with optional Laplace noise (simple, deterministic in tests when anonymize=False).
- How to Extend: add new adapters, contracts, or playbooks under respective namespaces; ensure tests cover new behavior.