pyproject.toml

[project]
name = "maestro"
version = "1.0.2"
description = "MAESTRO - Multi-Agent Evaluation for Structured Relational Output. Thesis artifact comparing orchestration frameworks for relational visualization."
readme = "README.md"
license = { text = "MIT" }
requires-python = ">=3.11,<3.12"
authors = [
    { name = "Colin Bolli", email = "colin.bolli@stud.fhgr.ch" }
]
keywords = ["multi-agent", "llm", "visualization", "orchestration", "thesis"]

dependencies = [
    # When adding a runtime-relevant dep here, also add it to _LIB_WHITELIST
    # in src/maestro/db/environment.py so its installed version is captured
    # in run_environments.lib_versions for every experiment invocation.
    # Pure dev tooling (pytest, ruff under [project.optional-dependencies])
    # is exempt: it does not affect experiment results.
    # Core
    "python-dotenv>=1.2.2",
    "pydantic>=2.13.3",
    "anthropic>=0.88",
    "openai>=2.30",
    # 2.4.6 was a compromised release (TanStack supply-chain incident); 2.4.7+
    # restructured the SDK layout: `Mistral` and `SDKError` moved out of the
    # package root. Pin >=2.4.7 to enforce both: skip-the-bad-release and
    # match-the-import-paths-this-code-uses.
    "mistralai>=2.4.7",
    "google-genai>=1.75.0",
    "crewai>=1.6.1",
    "langgraph>=1.1.10",
    # Provider-level retry with exponential backoff (see providers/_retry.py).
    # Already pulled in transitively by langgraph/crewai but declared
    # explicitly so it doesn't silently disappear on a transitive bump.
    # >=8.1.0: that release introduced `wait_exponential_jitter`, which
    # _retry.py imports directly. 8.0.x lacks the symbol.
    "tenacity>=9.1.4",
    # Statistical analysis pipeline (src/maestro/analysis/statistics.py).
    # statsmodels provides the factorial ANOVA (ols + anova_lm) and Tukey
    # HSD used for the thesis-grade analysis; pandas is its tabular input
    # layer (and statsmodels pulls it in regardless) so we depend on it
    # explicitly rather than relying on the transitive pin. scipy arrives
    # transitively via statsmodels and is not declared here directly.
    "statsmodels>=0.15.0",
    # pandas 3.0.4 was yanked from PyPI; 3.0.3 is the latest non-yanked release.
    "pandas>=3.0.5",
    # scipy is a transitive dependency of statsmodels (it supplies the
    # underlying distributions for the F-test, Tukey HSD, etc.), so its
    # version materially affects the statistics output. Not declared as a
    # direct dependency, but listed in _LIB_WHITELIST (db/environment.py) so
    # its installed version is captured in run_environments.lib_versions.
    # Visualizer (src/maestro/viz/): a Streamlit dashboard reading the
    # experiment DB read-only. Kept in core (not an extra) because this is a
    # single-image monorepo: the container serves the experiment OR the viz,
    # never both at once, and the extra footprint is negligible at the scale
    # Docker already handles. One install (pip install -e .) covers everything.
    #
    # Charts are rendered with matplotlib (displayed via st.pyplot), not a
    # browser-component library: the thesis deliverable is static,
    # publication-quality figures, and matplotlib gives one style system for
    # both the dashboard and the embedded thesis figures with no
    # browser-component fragility.
    "streamlit>=1.63.0,<2.0",
    "matplotlib>=3.11.1,<4",
    # Windows-only: provides the IANA timezone database that zoneinfo
    # (used by analysis/timestamps.py) needs at runtime. macOS and most
    # Linux distros ship the DB with the OS, so this is a no-op there:
    # the environment marker keeps the install footprint unchanged
    # on the development OS. Without this, a Windows reviewer running
    # the timestamp tests (or the analysis layer once #19/#24 ship)
    # would hit ZoneInfoNotFoundError on every named zone.
    "tzdata; sys_platform == 'win32'",
]

[project.optional-dependencies]
# Dev tooling - install with: pip install -e ".[dev]"
dev = [
    "pytest>=9.1.1",
    # Cap below 0.16: that release turned on formatting of Python inside
    # markdown fences by default, which reformats doc code blocks (and
    # misreads illustrative fragments, e.g. a trailing-comma snippet as a
    # tuple). Adopt it deliberately later with an explicit config, not via a
    # floating CI install. The pre-commit rev tracks this bound in lockstep.
    "ruff>=0.15.22,<0.16",
    # Strips notebook outputs/exec-counts/metadata; also wired as a pre-commit
    # hook. Here so the CLI is available to strip a notebook on demand.
    "nbstripout>=0.9.1",
]

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

[tool.hatch.build.targets.wheel]
packages = ["src/maestro"]

[tool.ruff]
line-length = 88
target-version = "py311"

[tool.ruff.lint]
select = ["E", "F", "I", "W"]

[tool.ruff.lint.per-file-ignores]
# run.py deliberately calls load_dotenv() and sets CrewAI env vars before
# importing strategy modules: those imports must see the configured
# environment, so they cannot move to the top of the file.
"src/maestro/run.py" = ["E402"]
# experiment_config.py is a data registry: each InputFile carries a one-line
# human description of the diagram. These are intentionally long prose strings
# where wrapping would hurt readability, so line length is not enforced here.
"src/maestro/experiment_config.py" = ["E501"]
# prompts.py holds the canonical Mermaid output contract: each rule is a single
# prose line shown to the model verbatim. Wrapping a rule would split it across
# bullets in the rendered prompt, so line length is not enforced here.
"src/maestro/prompts.py" = ["E501"]

[tool.pytest.ini_options]
testpaths = ["tests"]
pythonpath = ["src"]