pyproject.toml

[build-system]
requires = ["setuptools>=68.0", "wheel"]
build-backend = "setuptools.build_meta"

[project]
name = "autonet-computer"
version = "0.7.1"
description = "Decentralized AI alignment network with agent orchestration framework"
requires-python = ">=3.11"
license = "MIT"
readme = "README.md"
authors = [
    {name = "ATN Contributors"},
]
classifiers = [
    "Development Status :: 4 - Beta",
    "Intended Audience :: Developers",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Programming Language :: Python :: 3.13",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
]
dependencies = [
    # ATN core (agent framework + full-node participation). A plain
    # `pip install autonet-computer` gives a functional node: agent
    # orchestration, chain registration, and substrate/p2p close. Only the
    # native-training stack is optional (see [training]).
    "rich>=13.0",
    "pyyaml>=6.0",
    "httpx>=0.27",
    "mcp>=1.0",
    "websockets>=13.0",
    "python-dotenv>=1.0.0",
    "requests>=2.28.0",
    # Vault (agent secret access-control) — ships in core, gated at runtime by
    # ATN_WORKER_ISOLATION. Self-contained age-encrypted keystore (no HashiCorp).
    # pyrage = age encryption for the local secret store (vendored under
    # atn/_vendor/kevin). The broker itself uses only stdlib + ctypes.
    "pyrage>=1.1",
    # psutil = PID-reuse defense (create_time/cmdline) for the agent supervisor's
    # orphan recovery under ATN_WORKER_ISOLATION. Cross-platform; the Linux broker
    # _identity has a /proc fast path that does not need it, but the supervisor
    # does. Small, pure-ish, ships in core.
    "psutil>=5.9.0",
    # --- Full-node participation (register + gossip + substrate close) ---
    # These are on the LIVE product path, so a plain `pip install
    # autonet-computer` yields a fully functional node. Only genuinely
    # compute-heavy training deps stay optional (see [training] below).
    #
    # Blockchain: agent registration + on-chain service. atn/on_chain.py,
    # atn/runtime/agent_registry.py, nodes/common/crypto.py all lazy-import
    # web3/eth-account; eth-utils.keccak is used unguarded at module level in
    # nodes/common/mint_merkle.py. cryptography (gossip ed25519 signing) ships
    # transitively with these.
    "web3>=6.0.0",
    "eth-account>=0.10.0",
    "eth-utils>=2.0.0",
    # P2P: libp2p event gossip + federated close transport (nodes/common/p2p.py,
    # blob_resolver.py). multiaddr ships with libp2p. trio is the required
    # async runtime for the libp2p host. (miniupnpc is NOT imported anywhere in
    # code — UPnP defaults off — so it stays optional under [network] below.)
    "libp2p>=0.6.0",
    "trio>=0.22.0",
    # Substrate engine math: numpy is imported UNGUARDED at module level on the
    # live federated-close path — nodes/common/world_model_substrate/reconcile.py
    # does `from world_model.generalized import World`, whose __init__ pulls
    # equilibrate.py (`import numpy as np`). Also coverage.py / artifact_index.py
    # (daemon-local retrieval). Pure-Python wheels, cheap. Must be core.
    "numpy>=1.24.0",
]
# Note: the substrate runtime needs the world-model engine. It's
# vendored under nodes/common/world_model_substrate/_engine/ so
# `pip install autonet-computer` works without external git deps.
# Source of truth: github.com/autonet-code/world-model (master).

[project.urls]
Homepage = "https://autonet.computer"
Repository = "https://github.com/autonet-code/node"
Issues = "https://github.com/autonet-code/node/issues"

[project.optional-dependencies]
# Native training: JEPA / FedAvg / model-sharding stack. These are the only
# genuinely compute-heavy deps and are NOT on the live product path — every
# torch/transformers/mss import on atn/ + nodes/common/ (outside the shelved
# JEPA modules ml.py, jepa.py, vl_jepa.py, text_jepa.py) is behind a
# try/except or a lazy in-function import. scipy is not imported by any live
# module either; it lives here with the rest of the training stack.
training = [
    "torch>=2.0.0",
    "scipy>=1.10.0",
    # LLM backbone (open-weights encoder for JEPA)
    "transformers>=4.40.0",
    "accelerate>=0.28.0",
    # Screen capture (perception input for training/observation)
    "mss>=9.0.0",
]
# Back-compat alias. The full-node deps (web3, eth-account, eth-utils, libp2p,
# trio, numpy) moved to core so `pip install autonet-computer` is a working
# node — see the block in [project].dependencies. This extra now carries only
# the optional UPnP helper (never imported in code; UPnP defaults off) plus the
# training stack, so existing `autonet-computer[network]` instructions still
# resolve. Prefer `[training]` for the training stack going forward.
network = [
    "autonet-computer[training]",
    "miniupnpc>=2.0",
]
# Core voice: TTS output + audio mixing
voice = [
    "numpy>=1.24",
    "sounddevice>=0.4",
]
# Individual TTS backends
voice-kokoro = ["autonet-computer[voice]", "kokoro-onnx>=0.4"]
voice-edge   = ["autonet-computer[voice]", "edge-tts>=6.0", "miniaudio>=1.5"]
voice-11labs = ["autonet-computer[voice]", "requests>=2.28"]
# Push-to-talk: STT input via hotkey recording + Whisper transcription
voice-ptt = ["autonet-computer[voice]", "keyboard>=0.13", "faster-whisper>=1.0"]
# Everything voice-related
voice-full = [
    "autonet-computer[voice]",
    "autonet-computer[voice-kokoro]",
    "autonet-computer[voice-edge]",
    "autonet-computer[voice-ptt]",
]
# Development
dev = [
    "pytest>=7.0.0",
    "pytest-asyncio>=0.21.0",
    "black>=23.0.0",
    "mypy>=1.0.0",
]

[project.scripts]
atn = "atn.cli:main"
atn-vault-setup = "atn.vault_setup:main"

[tool.setuptools.packages.find]
include = ["atn*", "nodes*", "bridge", "world_model*"]

[tool.setuptools.package-data]
"atn.connectors" = ["*/connector.json"]
"bridge" = ["*.ts", "package.json", "bun.lock"]
"atn._vendor.kevin.vault" = ["*.md"]

[tool.pytest.ini_options]
asyncio_mode = "auto"
testpaths = ["tests"]