pyproject.toml

[build-system]
requires = ["setuptools>=69"]
build-backend = "setuptools.build_meta"

[project]
name = "whileai"
version = "0.131"
description = "whileai (import whileai as wai): post-training data and evaluation for tool-using LLM agents. Simulate, grade, measure, select, train, export, with the checks from rlhfbook.com."
readme = "README.md"
license = {text = "Apache-2.0"}
requires-python = ">=3.10"
dependencies = ["requests>=2.25.0", "pydantic>=2.0"]
authors = [{name = "Jacob Weiss", email = "jacob@while.ai"}]
keywords = [
    "wai", "while", "whileai", "evals", "agent-evals", "llm-evals", "llm-as-a-judge",
    "pass-at-k", "held-out", "confidence-interval", "post-training", "open-weights",
    "byok", "rlhf",
    "rl", "reinforcement-learning", "grpo", "sft", "dpo", "synthetic-data", "agents",
    "ai-agents", "tool-use", "evaluation", "simulation", "llm", "claude-code",
    "agents-md", "cursor", "codex",
]
classifiers = [
    "Development Status :: 3 - Alpha",
    "Intended Audience :: Developers",
    "Intended Audience :: Science/Research",
    "License :: OSI Approved :: Apache Software License",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.10",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Programming Language :: Python :: 3.13",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
    "Typing :: Typed",
]

[project.urls]
Homepage = "https://while.ai"
Documentation = "https://docs.while.ai"
Repository = "https://github.com/whilehq/whileai-sdk"
Issues = "https://github.com/whilehq/whileai-sdk/issues"
Datasets = "https://huggingface.co/while-ai"

[project.scripts]
# The command is `wai`: three letters, one token, the same name as the
# import alias. `whileai` runs the same entry point for anyone who typed
# the package name; `zp` was retired on 2026-09-21.
wai = "whileai.cli:main"
whileai = "whileai.cli:main"
whileai-simulations = "whileai.simulations.score.quality:main"

[project.optional-dependencies]
dev = ["pytest>=8.0", "pytest-cov>=5.0", "pytest-xdist>=3.6", "ruff>=0.12",
       "mypy>=1.16", "types-requests", "ty>=0.0.82", "math-verify>=0.8"]
rl = ["verifiers>=0.3; python_version >= '3.11' and python_version < '3.14'"]
# export_environment(runtime="openenv") needs nothing but the SDK; serving the
# package (whileai.simulations.openenv) needs Meta PyTorch's OpenEnv, which
# brings fastapi, fastmcp and gradio, so it stays an extra.
openenv = ["openenv>=0.5"]
# verify.MathEqual decides with Math-Verify (Kydlicek et al. 2025); imported
# in the constructor, which names this extra when it is missing. No fallback:
# a weaker rule is a reward the policy learns to game. Optional so the base
# install stays two deps; dev carries it so the tests run.
math = ["math-verify>=0.8"]
# wai.hub.push / export(push_to=): the Hub client that authenticates with the
# user's own HF_TOKEN; imported lazily, so the base install stays two deps.
hf = ["huggingface_hub>=0.20"]
# wai.models.Bedrock on AWS credentials (aws configure, AWS_PROFILE, a role): boto3
# signs the Converse calls. A Bedrock API key in AWS_BEARER_TOKEN_BEDROCK
# needs nothing here; the calls then go over requests.
bedrock = ["boto3>=1.35"]

[tool.setuptools]
include-package-data = false

[tool.setuptools.package-data]
whileai = ["py.typed"]
"whileai.simulations" = ["py.typed", "schemas/*.json"]

[tool.setuptools.packages.find]
where = ["."]
include = ["whileai*"]
# behaviors is internal tooling for our own training runs: in the repo,
# never in the wheel.
exclude = ["tests*", "recipes*", "docs*", "specs*",
           "whileai.simulations.behaviors*"]
namespaces = false

[tool.pytest.ini_options]
testpaths = ["tests"]
pythonpath = ["."]
# One worker per core. The suite is 1,600 small tests with no shared state;
# serial it takes about three minutes, on four cores about forty seconds.
# Pass -n0 for a serial run (needed for --pdb and for reading -s output in order).
addopts = "-n auto"

[tool.ruff]
line-length = 100
target-version = "py310"
extend-exclude = ["build", "dist"]

[tool.ruff.lint]
# PLR2004 (magic value in a comparison) is the ruff half of the
# no-hardcoding rule: scripts/check_no_hardcoding.py checks the comment
# form on named constants, PLR2004 catches the bare number in a compare
# whatever shape it sits in. Strings and bytes are exempt (status words,
# role names); numbers are not, so a `2` that is structure carries a
# `# noqa: PLR2004  # <why>`.
select = ["E", "W", "F", "I", "UP", "B", "SIM", "RUF", "PLR2004"]
ignore = [
    "E501",   # long lines: the formatter owns wrapping; prose in docstrings stays
    "RUF001", "RUF002", "RUF003",  # typographic dashes and arrows in prose
    "RUF012",  # mutable class defaults are used deliberately as registries
    "SIM115",  # open() without a context manager where the handle is returned
    "B905",    # zip(strict=) everywhere reads worse than the cases it protects
]

[tool.ruff.lint.pylint]
allow-magic-value-types = ["str", "bytes"]

[tool.ruff.lint.per-file-ignores]
# defaults.py documents every constant as `# NAME = value: why`, which
# ERA001 (commented-out code) would read as dead code if it is ever enabled.
"whileai/simulations/defaults.py" = ["ERA001"]
# The package init re-exports the public surface; the re-exports are the point.
"whileai/simulations/__init__.py" = ["F401", "E402"]
"whileai/__init__.py" = ["F401"]
# Examples and the identity generator set sys.path before importing.
"recipes/**" = ["E402", "PLR2004"]
"tests/**" = ["E402", "PLR2004"]
"scripts/**" = ["PLR2004"]
# Skill playbooks are recipes: their check.py thresholds are the lesson, not engine defaults.
"skills/**" = ["PLR2004"]
".github/**" = ["PLR2004"]
# The account, ingest and platform clients outside simulations/ are not
# under the no-hardcoding rule (yet); their numbers are HTTP codes and
# gate limits. Named one by one: a `whileai/*.py` glob would match the
# whole package.
"whileai/auth.py" = ["PLR2004"]
"whileai/init_evals.py" = ["PLR2004"]
"whileai/platform.py" = ["PLR2004"]
"whileai/cli.py" = ["PLR2004"]

[tool.ruff.lint.isort]
known-first-party = ["whileai", "tests"]

[tool.mypy]
python_version = "3.10"
packages = ["whileai"]
ignore_missing_imports = true
check_untyped_defs = true
warn_unused_ignores = true
warn_redundant_casts = true
no_implicit_optional = true

# Debt: these modules hang attributes on function objects at runtime
# (the situation writer is a closure with state). They type-check once
# the writer is a class. Tracked in CHANGELOG under hygiene.
[[tool.mypy.overrides]]
module = [
    "whileai.simulations.generate.generator",
    "whileai.simulations.generate.scenarios",
    "whileai.simulations.run.engine",
]
ignore_errors = true

# ty runs beside mypy in CI (about ten times faster cold); mypy stays the
# gate until ty reaches 1.0. Same shape as the mypy block above: optional
# extras (torch, sympy, verifiers, ...) import lazily inside functions, so
# an unresolved import is an extra that is not installed, not a typo.
[tool.ty.src]
include = ["whileai"]

[tool.ty.environment]
python-version = "3.10"

[tool.ty.rules]
unresolved-import = "ignore"

# The same three modules as the mypy override: attributes hung on function
# objects. Drop this with the mypy override once the writer is a class.
[[tool.ty.overrides]]
include = [
    "whileai/simulations/generate/generator.py",
    "whileai/simulations/generate/scenarios.py",
    "whileai/simulations/run/engine.py",
]
[tool.ty.overrides.rules]
unresolved-attribute = "ignore"
invalid-argument-type = "ignore"
invalid-assignment = "ignore"
call-non-callable = "ignore"
not-subscriptable = "ignore"