pyproject.toml

[build-system]
requires = [
    "setuptools >= 84.0.0",
    "setuptools-scm >= 8",
]
build-backend = "setuptools.build_meta"

[project]
# The distribution is `reef-infra`; the import package remains `reef`.
name = "reef-infra"
dynamic = ["version"]
description = "A framework-agnostic runtime for continuously improving agents"
readme = "README.md"
license = "Apache-2.0"
authors = [{ name = "Human-Agent-Society" }]
requires-python = ">=3.12"
classifiers = [
    "Programming Language :: Python :: 3.12",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
    "Topic :: System :: Distributed Computing",
]
# Reef core dependencies. Keep this list free of training and GPU packages:
# `pip install reef-infra` must support the base `recipe` configuration on CPU.
dependencies = [
    "aiohttp>=3.14",  # web.RequestKey
    "alembic>=1.13,<2",
    "huggingface_hub",
    # Harbor's environment contract backs reef.execution.sandbox; task directory rollouts, the task player and the
    # generator service play Harbor tasks with reef-eval.
    "harbor>=0.20,<0.25",
    "pyyaml",
    "reef-client>=0.2.0",
    "reef-eval>=0.1.1",
    "sqlalchemy>=2.0,<3",
    "tomli-w",
]

[project.optional-dependencies]
# Python-side inference adapters; install native SGLang in its GPU environment.
sglang = ["httpx", "requests", "ray[default]", "sglang-router>=0.3.2", "psutil>=5.9"]
# Python-side vLLM adapter; install native vLLM in its GPU environment.
vllm = ["httpx", "requests", "ray[default]", "psutil>=5.9"]
tinker = [
    "tinker==0.35.0",
    # Converts Tinker checkpoints into PEFT adapters for a local engine (pulls torch).
    "tinker-cookbook==0.5.8",
    "jinja2>=3.1.0",
]
postgres = ["psycopg[binary]>=3.1,<4"]
# Reads s3:// sources of an artifact import (storage.artifact_sources).
s3 = ["boto3"]
# Cloud sandboxes for evaluation rollouts and the Reefine agent: E2B, and Modal. Harbor's extra of each
# environment adds what its sandbox needs to build a task's image (dockerfile-parse).
e2b = ["harbor[e2b]", "e2b>=2.46", "connectrpc>=0.11.1,<0.13"]
modal = ["harbor[modal]", "modal>=1.6"]
# Provider-neutral experiment tracking. Non-Slime training backends can opt in
# without installing the Slime/GPU dependency stack.
wandb = [
    "wandb>=0.29.0",
]
# Record tracing over OTLP/HTTP (observability.tracing); any OpenTelemetry-compatible backend receives it.
opentelemetry = [
    "opentelemetry-sdk>=1.25",
    "opentelemetry-exporter-otlp-proto-http>=1.25",
]
# Python-side dependencies used by the Slime backend. The runtime source pin
# lives in the `runtime` dependency group below.
# torch / SGLang / Megatron-Core / FlashAttention remain supplied by the
# supported slimerl/slime GPU image because their versions are CUDA-specific.
slime = [
    "e2b",
    "httpx[http2]",
    "memray",
    "omegaconf",
    "peft",
    "pillow",
    "psutil>=5.9",
    "pylatexenc",
    "qwen_vl_utils",
    "requests",
    "ray[default]",
    "safetensors",
    "sglang-router>=0.3.2",
    "tensorboard",
    "transformers",
    "wandb>=0.29.0",
]
# Local dev toolchain: `pip install -e .[dev]`. Versions of the lint
# tools stay in sync with .pre-commit-config.yaml revs.
dev = [
    "mypy",
    "pre-commit",
    "pytest",
    # CI measures coverage against the [tool.coverage.report] floor below.
    "pytest-cov",
    # Runs the suite in parallel, as CI does (-n 8; addopts selects --dist worksteal).
    "pytest-xdist==3.8.0",
    "ruff==0.17.2",
]

[dependency-groups]
# Source-installed runtime pins used by development, CI, and images. Keeping
# them out of package metadata lets GPU environments install them with
# `--no-deps` without replacing their CUDA-matched dependency stack.
runtime = [
    "slime @ git+https://github.com/THUDM/slime.git@540ea919e45b8b929198abce008a103be9cfc779",
]

[project.urls]
Homepage = "https://github.com/Human-Agent-Society/reef"

[project.scripts]
reef = "reef.cli:main"
reef-native = "reef.harness.runners.native.__main__:main"

[project.entry-points."sglang.srt.plugins"]
reef = "reef.inference.sglang.plugin:install_sglang_plugin"

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

[tool.setuptools_scm]
fallback_version = "0.0.0.dev0"

# reef-client is now a separate repo (Human-Agent-Society/reef-client) and a
# declared dependency; the reef wheel must not ship its source.
# `reef` is the shared machinery. The sibling `recipes` tree is a cookbook,
# not package code, so neither its methods nor its runnable examples ship.
# runtime/ intentionally contains only its five modules and executor/ (PEP 420).
[tool.setuptools.packages.find]
namespaces = true
include = ["reef", "reef.*"]

[tool.setuptools.package-data]
reef = [
    "harness/adapters/*/adapter.yaml",
    "harness/adapters/codex/default_instructions.md",
    "harness/adapters/codex/bundled_models.json",
    "train/harness_backend/agent/protocol.md",
]

[tool.isort]
profile = "black"  # black-compatible
line_length = 119  # should match black parameters
ignore_whitespace = true  # ignore whitespace for compatibility with the initial style
py_version = 312  # sync with requires-python in [project]
sections = ["FUTURE", "STDLIB", "THIRDPARTY", "FIRSTPARTY", "LOCALFOLDER"]
default_section = "THIRDPARTY"
known_first_party = ["reef", "recipes"]
known_third_party = ["megatron", "wandb", "ray", "transformers"]
src_paths = ["reef", "recipes"]
# Files no formatter touches (vendored benchmark content, saved result programs)
# are listed once, in .pre-commit-config.yaml's top-level `exclude`.


[tool.black]
line_length = 119

[tool.ruff]
target-version = "py312"  # sync with requires-python in [project]
line-length = 119  # sync with [tool.black] and [tool.isort]

[tool.ruff.lint]
select = [
    "E", "W",  # pycodestyle
    "F",       # pyflakes
    "B",       # flake8-bugbear
    "UP",      # pyupgrade
    "C4",      # flake8-comprehensions
    "SIM",     # flake8-simplify
    "PIE",     # flake8-pie
    "ISC",     # flake8-implicit-str-concat
    "RSE",     # flake8-raise
    "RET",     # flake8-return
    "PERF",    # perflint
    "LOG",     # flake8-logging
    "N",       # pep8-naming
    "RUF",     # ruff
]
ignore = [
    # E402: module-import-not-at-top-of-file
    "E402",
    # E501: line too long - black handles wrapping; long strings and URLs stay readable
    "E501",
    # N818: existing public exception names omit an Error suffix; renaming them would break callers
    "N818",
    # UP046/UP047: PEP 695 type parameter syntax; the codebase declares TypeVars, and mypy reads both forms
    "UP046",
    "UP047",
]
# Math notation in the RL objective docstrings: SAO's alpha/gamma, and the
# true minus sign used alongside other Unicode math (∇, π, Â) in the SAO loss
# derivation. Other Greek symbols (theta, pi, epsilon, lambda, delta) are not
# flagged; these three only trip RUF002/003 because they have ASCII lookalikes.
allowed-confusables = ["α", "γ", "−"]

[tool.ruff.lint.per-file-ignores]
# Test code: perf lint and strict class/branch shape rules add churn, not safety.
"tests/**/*.py" = ["N806", "PERF", "RUF012", "SIM108"]
# The Top-K objective follows the mathematical notation used in its derivation.
"recipes/openclawrl/slime/objective.py" = ["N802", "N803", "N806"]

[tool.mypy]
# Reef's training adapters are checked with the rest of the package; the
# runtime remains a separately installed distribution.
files = ["reef"]
python_version = "3.12"  # sync with requires-python in [project]
ignore_missing_imports = true

# The training runtime and its GPU dependencies do not ship type information
# with Reef.
[[tool.mypy.overrides]]
module = ["slime.*", "slime_plugins.*", "vllm.*"]
ignore_missing_imports = true

[tool.coverage.run]
source = ["reef"]
# Coverage from the GitHub and Blacksmith checkouts is combined in one gate.
relative_files = true
# Reef-authored training control flow counts toward the CPU coverage gate. These
# three modules execute only inside a live Megatron/CUDA worker; their useful
# validation is the GPU integration suite, not import-only CPU mocks.
omit = [
    "reef/train/slime_backend/reef_adapters/megatron/hf_export.py",
    "reef/train/slime_backend/reef_adapters/megatron/model_provider.py",
    "reef/train/slime_backend/reef_adapters/megatron/train_actor.py",
]

[tool.coverage.report]
skip_empty = true
precision = 1
# Floor just under the 82 percent baseline: prevents silent decay while the
# team decides a real target.
fail_under = 80

[tool.pytest.ini_options]
# Report the 20 slowest tests; --durations=0 lists every test.
# With -n, worksteal spreads single tests over the workers and moves queued
# tests from a busy worker to an idle one, so one slow file cannot hold the run.
# Every test must therefore pass alone and in any order. Without -n, --dist has no effect.
addopts = "--verbose --pyargs --durations=20 --strict-markers --import-mode=importlib --dist worksteal"
testpaths = ["./tests"]  # must be an explicit path to avoid importing another "tests" module
# directories to ignore when discovering tests
norecursedirs = [
    "examples",
    "*.egg",
    ".*",
    "_darcs",
    "build",
    "CVS",
    "dist",
    "venv",
    "{arch}",
]
# markers to select tests, use `pytest --markers` to see all available markers, `pytest -m "<marker>"` to select tests
markers = [
    "sandbox: requires real nested bubblewrap jails on the GitHub CI runner",
    "unit: marks unit test, i.e. testing a single, well isolated functionality (deselect with '-m \"not unit\"')",
    "integration: marks test checking the elements when integrated into subsystems (deselect with '-m \"not integration\"')",
    "acceptance: marks test checking whether the developed product/model passes the user defined acceptance criteria (deselect with '-m \"not acceptance\"')",
]