Skip to content

AGENTS.md

Latest commit

 

History

History
14 lines (12 loc) · 1016 Bytes

AGENTS.md

File metadata and controls

14 lines (12 loc) · 1016 Bytes
Raw

AGENTS.md — python/

Adds to the root AGENTS.md; never overrides it.

  • nftbench is installed editable (pip install --no-build-isolation -e .); tests live in python/tests/ and run with <venv-python> -m pytest -q from the repository root.
  • Write the failing test first. Tests that drive nft_export/nft_run need the native build; they fail (never skip) without it.
  • The scorer depends on no engine. schema.py, geometry.py, metrics.py, score.py read only banks and results. A new engine needs a runner, not scorer changes.
  • Metric definitions are part of the published results: changing one (e.g. what median_px covers) changes every table. Say so in the PR and the changelog, and regenerate the tables.
  • segment.py produces ground truth for real footage: it must return no corners rather than wrong ones. Read colour frames from the source video, not the grey bank.
  • publish.py holds the sensitive-data scan; extend its patterns rather than bypassing it.