Skip to content

AGENTS.md

Latest commit

 

History

History
40 lines (34 loc) · 1.87 KB

AGENTS.md

File metadata and controls

40 lines (34 loc) · 1.87 KB
Raw

Agent Instructions

Package Manager

Use pip with requirements.txt (Python 3.11). No lock file — pin new deps with >= at minimum.

Commit Attribution

AI commits MUST include:

Co-Authored-By: (the agent model's name and attribution byline)

File-Scoped Commands

Task Command
Run single test file python -m pytest backend/tests/test_rules_engine.py -v
Run one test python -m pytest backend/tests/test_classify.py::test_cdss_classified_as_high_risk
Full suite python -m pytest backend/tests/
Start backend python -m backend.main
Start Streamlit streamlit run app/main.py

Key Conventions

  • Never downgrade a risk classification — rules_engine.py escalates only; if in doubt, go higher.
  • All regulatory text lives in backend/data/*.json — never hardcode article text in Python.
  • Pydantic everywhere — no raw dicts in API responses. Models in backend/models/.
  • Stub the Claude agent in tests — patch backend.routers.classify._get_agent; never hit the real API in CI.
  • SaMD flag is mandatory for anything that could qualify as a medical device under MDR/IVDR.
  • PDF template at templates/audit_report.html uses Jinja2 + WeasyPrint; needs Pango/Cairo system libs.

Architecture

  • backend/services/classification_agent.py — Claude wrapper (structured JSON out)
  • backend/services/rules_engine.py — static validator that escalates
  • backend/routers/ — FastAPI endpoints (/classify, /report, /health, /examples)
  • app/main.py — Streamlit entry; launches FastAPI as subprocess on :8000
  • HF Spaces deploy: see docs/DEPLOYMENT.md

Coding Standards

  • Type hints on every function; docstrings with Args/Returns on public ones
  • Functions ≤ 50 lines
  • logging, never print()
  • No emojis in code (UI or otherwise) except the HF frontmatter 🏥