pyproject.toml

[build-system]
requires = ["setuptools>=68"]
build-backend = "setuptools.build_meta"

[project]
name = "parbench"
version = "0.1.0"
description = "Benchmark suite for evaluating LLM-based parallel code translation"
requires-python = ">=3.12"
dependencies = [
    "pydantic>=2.0",
    "jsonschema>=4.20",
    "libclang>=18.1",
]

[project.optional-dependencies]
eval = [
    "anthropic>=0.40.0",
    "openai>=1.50.0",
]
analysis = [
    "matplotlib>=3.9",
    "numpy>=1.26",
    "scipy>=1.14",
    "scienceplots>=2.2",
]
dev = [
    "pytest>=8.0",
    "ruff>=0.6.0",
]
all = [
    "parbench[eval,analysis,dev]",
]

[tool.ruff]
# line-length feeds `ruff format` only; E501 is deliberately NOT selected
# (decision D8, 2026-08-15): most lines sit under 100 and lighting up long
# lines across untouched files would flood the pre-existing-debt WARN bucket.
line-length = 100
target-version = "py312"
# Benchmark trees are never linted or fixed (CLAUDE.md: never edit benchmark
# source). HeCBench-master and the three *-src trees are already gitignored;
# rodinia is a submodule, so it needs an explicit exclusion.
extend-exclude = [
    "rodinia",
    "HeCBench-master",
    "mixbench",
    "xsbench",
    "rsbench",
]

[tool.ruff.lint]
# Explicit selection (rule 38 redesign, 2026-08-15): the classic core set plus
# import order, pyupgrade, and bugbear. Pinned explicitly because ruff's
# built-in defaults widened around 0.15 and a version bump silently changed
# which rules health_check's lint gate enforced.
select = ["E4", "E7", "E9", "F", "I", "UP", "B"]

[tool.pytest.ini_options]
markers = [
    "integration: requires real benchmark files on disk (GPU machine only)",
    "llm: real LLM API calls (gated by PARBENCH_RUN_LLM_TESTS=1; costs ~$5.81 for full smoke set)",
]

[tool.setuptools.packages.find]
where = ["."]
include = ["c_augmentation*", "harness*"]