pyproject.toml

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

[project]
name = "gym-anything"
version = "0.1.0"
description = "A unified runtime and benchmark suite for computer-use agent research."
readme = "README.md"
requires-python = ">=3.10"
license = {text = "MIT"}
authors = [
    {name = "Pranjal Aggarwal"},
]
classifiers = [
    "Development Status :: 3 - Alpha",
    "Intended Audience :: Science/Research",
    "Intended Audience :: Developers",
    "License :: OSI Approved :: MIT License",
    "Operating System :: OS Independent",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.10",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
]
dependencies = [
    "dbus-next>=0.2.3",
    "jsonschema>=4.20",
    "numpy>=1.26",
    "paramiko>=3.4",
    "Pillow>=10.0",
    "pycryptodome>=3.20",
    "PyYAML>=6.0",
    "requests>=2.31",
    "rich>=13.0",
]

[project.optional-dependencies]
# Everything the benchmark corpus' verifier scripts import. Mirrors
# requirements_benchmark.txt (generated by scripts/extract_benchmark_deps.py);
# keep the two in sync.
# Lower bounds are deliberate: with ~25 unbounded packages here, pip's
# resolver dies with `resolution-too-deep` backtracking through years of
# old releases (the floors prune that search space; uv resolves either
# way). Keep every floor at or below the currently resolving version set.
benchmark = [
    # astropy 7+ drops Python 3.10, which this project still supports
    "astropy>=6.1",
    "beautifulsoup4>=4.12",
    "biopython>=1.84",
    "dbfread>=2.0",
    "ezdxf>=1.3",
    "fuzzywuzzy>=0.18",
    "gimpformats>=2023",
    "lxml>=5.3",
    "nibabel>=5.2",
    "numpy>=1.26",
    "odfpy>=1.4",
    "openai>=1.60",
    "opencv-python>=4.10",
    "openpyxl>=3.1",
    "pandas>=2.2",
    "pdfminer.six>=20240706",
    "phpserialize>=1.3",
    "Pillow>=10.0",
    "pynrrd>=1.0",
    "pypdf>=5.0",
    "PyPDF2>=3.0",
    "pyshp>=2.3",
    "python-dateutil>=2.9",
    "python-docx>=1.1",
    "python-Levenshtein>=0.25",
    "python-pptx>=1.0",
    "pytz>=2024.1",
    "PyYAML>=6.0",
    "requests>=2.31",
    "scikit-image>=0.24",
    "scikit-learn>=1.5",
    "scipy>=1.13",
    "shapely>=2.0",
    "statsmodels>=0.14",
    # Transitive dep of scikit-image, named here only to bound the resolver
    # (it was the top backtracking hotspot at 66 candidate visits).
    "tifffile>=2025.1.10",
    "trimesh>=4.10",
    "vitaldb>=1.7",
]
agents = [
    "anthropic>=0.50",
    "backoff>=2.2",
    # 2.x adds ENVIRONMENT_DESKTOP/_MOBILE, which GeminiComputerUseAgent
    # uses for gemini-3.5-flash (older SDKs fall back to browser).
    "google-genai>=2.10",
    "litellm>=1.52",
    "loguru>=0.7",
    # 1.66 adds the Responses API, which GPT54ComputerUseAgent requires
    "openai>=1.66",
    "python-dotenv>=1.0",
    "tqdm>=4.66",
]
services = [
    "Flask>=3.0",
    "flask-cors>=4.0",
    "httpx>=0.27",
    "hypercorn>=0.17",
    "psutil>=5.9",
    "quart>=0.19",
    "quart-cors>=0.7",
    "requests>=2.31",
]
vlm = [
    "anthropic>=0.50",
    "litellm>=1.52",
    "openai>=1.0",
]
use_computer = [
    # Floor at the release the runner surface was last validated against:
    # top-level MacOSSandbox import, exec_ssh/upload_dir/download_dir,
    # mouse/keyboard/screenshot signatures.
    "use-computer>=0.0.44",
]
modal = [
    "modal>=1.0",
]
modal_native = [
    "modal>=1.4",
]
prime-rl = [
    "datasets>=4.0",
    # Transitive dep of datasets/tokenizers/verifiers, named here only to
    # bound the resolver (51 candidate visits without a floor).
    "huggingface-hub>=1.0",
    "verifiers>=0.1.14",
    # The prime-rl rollout drives a real reference agent (agents.agents), whose
    # registry import pulls in every agent class (e.g. litellm via ClaudeAgent),
    # so gym-anything[prime-rl] must carry the agents runtime deps or the env
    # fails to load with ModuleNotFoundError.
    "gym-anything[agents]",
]
all = [
    "gym-anything[benchmark,agents,services,vlm,use_computer,modal,modal_native,prime-rl]",
]

[project.urls]
Homepage = "https://cmu-l3.github.io/gym-anything/"
Documentation = "https://cmu-l3.github.io/gym-anything/docs/"
Repository = "https://github.com/cmu-l3/gym-anything"
Issues = "https://github.com/cmu-l3/gym-anything/issues"

[project.scripts]
gym-anything = "gym_anything.cli:main"
gym-anything-dashboard = "gym_anything.remote.dashboard_app:main"
gym-anything-extras = "gym_anything.extras_cli:main"
gym-anything-master = "gym_anything.remote.master:main"
gym-anything-trajectory-viewer = "gym_anything.dashboards.trajectory.app:main"
gym-anything-worker = "gym_anything.remote.worker:main"

[tool.setuptools]
include-package-data = true
package-dir = {"" = ".", "gym_anything" = "src/gym_anything"}
packages = [
    "gym_anything",
    "gym_anything.config",
    "gym_anything.dashboards",
    "gym_anything.dashboards.trajectory",
    "gym_anything.integrations",
    "gym_anything.integrations.harbor",
    "gym_anything.integrations.prime_rl",
    "gym_anything.presets",
    "gym_anything.recording",
    "gym_anything.remote",
    "gym_anything.remote.monitoring",
    "gym_anything.runners",
    "gym_anything.runtime",
    "gym_anything.runtime.recording",
    "gym_anything.runtime.runners",
    "gym_anything.tui",
    "gym_anything.utils",
    "gym_anything.verification",
    "agents",
    "agents.evaluation",
    "agents.agents",
    "agents.shared",
    "benchmarks",
    "benchmarks.cua_world",
    "benchmarks.cua_world.registry",
]

[tool.setuptools.package-data]
gym_anything = [
    "dashboards/trajectory/templates/*.html",
    "presets/*.json",
    "presets/**/*.json",
    "presets/**/*.sh",
    "presets/**/Dockerfile",
    "remote/monitoring/templates/*.html",
    "runtime/runners/*.def",
    "runtime/runners/windows_setup/*.py",
    "runtime/runners/modal_native_assets/*",
]
benchmarks = [
    "cua_world/splits/*.json",
    "cua_world/reports/*.json",
    "cua_world/reports/*.txt",
    # The benchmark content itself, so installed copies (git/hub installs)
    # can run environments without a repo checkout.
    "cua_world/environments/**/*",
    "cua_world/environments/**/.*",
]

[tool.setuptools.exclude-package-data]
"*" = ["*.pyc", "*.pyo", "**/__pycache__/*"]