pyproject.toml

[project]
name = "daedalus"
version = "1.0.0"
description = "DAEDALUS: bootstrapping agent memory from self-generated tasks (AppWorld, tau2-bench, AutomationBench)"
readme = "README.md"
license = "MIT"
license-files = ["LICENSE"]
requires-python = ">=3.12"
dependencies = [
    "litellm>=1.55",      # one call surface for every model provider
    "pyyaml>=6.0",
    "jinja2>=3.1",
    "rank-bm25>=0.2.2",   # BM25 retriever
    "tqdm>=4.66",
    "tiktoken>=0.7",      # token counts for the generation-cost breakdown of pre-ledger runs
]

[project.optional-dependencies]
dev = ["pytest", "ruff"]
plots = ["matplotlib>=3.8", "scipy>=1.11"]
# Dense retrieval (Qwen3-Embedding), lazy-imported in core/memory/retrieval.py.
dense = ["sentence-transformers>=3.0", "faiss-cpu>=1.7"]

# Benchmarks. AppWorld and tau2-bench are pinned as git references at the commits the
# reported runs used (neither name resolves to the right project on PyPI);
# [tool.uv.sources] below redirects both to the editable benchmarks/ submodules, because
# tau2 reads its data/ from the source tree.
appworld = [
    "appworld @ git+https://github.com/StonyBrookNLP/appworld@a072b7a86e7c1d5b1d7175659d750ebb9b79f10a",
]
tau2 = [
    "tau2 @ git+https://github.com/sierra-research/tau2-bench@1901a301961cbbe3fd11f3e84a2a376530c759e3",
    "audioop-lts>=0.2.1; python_version >= '3.13'",
]
# AutomationBench is read from the benchmarks/AutomationBench submodule (on sys.path) rather
# than installed, which would pull its unused `verifiers` runner; its task builders need
# `datasets`.
automationbench = ["datasets>=5.0.0"]

all-benchmarks = ["daedalus[appworld,tau2,automationbench,dense,plots]"]

[tool.uv.sources]
appworld = { path = "benchmarks/appworld", editable = true }
tau2 = { path = "benchmarks/tau2-bench", editable = true }

[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"

[tool.hatch.build.targets.wheel]
packages = ["daedalus"]

# The appworld/tau2 extras are pinned git URLs, which hatchling rejects by default.
[tool.hatch.metadata]
allow-direct-references = true

[tool.pytest.ini_options]
# Offline unit tests only. The smoke tests (tests/smoke_*.py) make real, paid LLM calls
# and need the benchmark checkouts; run them explicitly, e.g. `uv run pytest tests/smoke_tau2.py`.
testpaths = ["tests"]
python_files = ["test_*.py"]

[tool.ruff]
line-length = 100
extend-exclude = ["outputs", "blogpost", "./benchmarks"]

[tool.ruff.lint]
select = ["E4", "E7", "E9", "F"]