pyproject.toml

[project]
name = "face-recognition-cli"
version = "0.8.0"
description = "Face recognition and identity management: detect faces, collect embeddings per identity, and manage them by generated id and human name — enroll, match, list, forget one, and forget all at once. Extracted from reachy-mini-cli's OpenCV YuNet+SFace engine so Reachy Mini depends on it instead of carrying its own copy."
readme = "README.md"
license = "Apache-2.0"
requires-python = ">=3.12"
authors = [{name = "AgentCulture"}]
classifiers = [
    "Development Status :: 3 - Alpha",
    "Programming Language :: Python :: 3.12",
    "License :: OSI Approved :: Apache Software License",
    "Topic :: Software Development",
    "Intended Audience :: Developers",
]
dependencies = ["numpy>=1.26,<3"]

# `numpy` is the only base dependency: both the ported FaceEngine (embedding
# arrays) and FaceStore (per-embedding .npy files) need it, so it can't stay
# behind an extra the way the template's zero-runtime-deps property would
# prefer — this deliberately gives that property up (issue #1, q1).
#
# [cpu] and [gpu] are compute-class extras, not model-class extras: both pull
# the SAME opencv-python-headless wheel and run the SAME ONNX models
# (YuNet detector + SFace embedder), producing the SAME embedding space. A
# face enrolled under one extra must match under the other — accelerating
# *execution* must never fork the model.
[project.optional-dependencies]
# The Raspberry-Pi-class robot box: the default path. headless because a
# robot box has no GUI to bind against. cv2 is lazy-imported inside
# face_recognition_cli's engine module only — never at import time — so a
# bare install without this extra stays importable and fails with a clean
# exit-2 CliError naming [cpu]/[gpu] instead of an ImportError traceback.
cpu = ["opencv-python-headless>=4.9,<5"]
# DGX Spark / Jetson / RTX-class hosts, for throughput. Same package, same
# models: v1 GPU acceleration is cv2's own CUDA DNN backend, used when the
# locally installed OpenCV build carries it (Jetson/DGX builds typically do;
# the PyPI opencv-python-headless wheel does not ship CUDA DNN, so this extra
# alone doesn't guarantee acceleration — the build backing it does). No
# second inference stack is introduced here, so the embedding space stays
# identical by construction. onnxruntime-gpu over the same ONNX files is the
# documented future path if throughput demands more than the DNN backend
# gives.
gpu = ["opencv-python-headless>=4.9,<5"]

[project.urls]
Homepage = "https://github.com/agentculture/face-recognition-cli"
Issues = "https://github.com/agentculture/face-recognition-cli/issues"

[project.scripts]
face-recognition = "face_recognition_cli.cli:main"

[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"

[tool.hatch.build.targets.wheel]
packages = ["face_recognition_cli"]

[dependency-groups]
dev = [
    "pytest>=8.0",
    "pytest-xdist>=3.0",
    "pytest-cov>=4.1",
    "bandit>=1.7.5",
    "flake8>=6.1",
    "isort>=5.12.0",
    "black>=23.7.0",
    "teken>=0.8",
]

[tool.coverage.run]
source = ["face_recognition_cli"]
omit = ["face_recognition_cli/__pycache__/*"]
# Emit repo-relative paths in coverage.xml so SonarCloud maps them to
# sonar.sources=face_recognition_cli (absolute/.venv paths fail to map → empty coverage).
relative_files = true

[tool.coverage.report]
fail_under = 60
show_missing = true
exclude_lines = [
    "pragma: no cover",
    "if __name__ == .__main__.",
    "if TYPE_CHECKING:",
]

[tool.isort]
profile = "black"
line_length = 100
known_first_party = ["face_recognition_cli"]

[tool.black]
line-length = 100
target-version = ["py312"]

[tool.bandit]
exclude_dirs = ["tests"]
skips = ["B101", "B404", "B603"]

[tool.pytest.ini_options]
testpaths = ["tests"]
addopts = "-ra"