[build-system]
requires = ["setuptools>=68", "setuptools-scm>=8", "wheel"]
build-backend = "setuptools.build_meta"

[project]
name = "strands-decider"
dynamic = ["version"]
description = "An experimental typed classifier: calibrated yes/no, choice and score answers from an LLM torso with a pointer readout."
readme = "README.md"
license = "Apache-2.0"
license-files = ["LICENSE"]
requires-python = ">=3.10"
dependencies = [
    "torch>=2.7",
    "transformers>=5.15,<6",
    "peft>=0.21",
    "accelerate>=0.34",
    "huggingface_hub>=1.5",
    "pydantic>=2.7",
    "typer>=0.12",
    "rich>=13.7",
    "pyyaml>=6.0",
    "fastapi>=0.115",
    "uvicorn[standard]>=0.30",
    "numpy>=1.24",
    "safetensors>=0.4",
]

[project.optional-dependencies]
# The corpus build and multistep converter use datasets:
# `python -m training.data build` and `python -m training.data.multistep`. Those modules
# live under training/, not src/, so they ship in the repo, not in the wheel; this extra
# is here for `pip install -e ".[train]"` from a checkout rather than for Hub consumers.
train = ["datasets>=2.20"]
# Image input (`serve --vision`, `ask --image`, docs/vision.md); also needs transformers >= 5.18.
vision = ["pillow>=10"]
# `--device mlx`: the torso on Apple silicon's GPU through mlx-lm (src/strands_decider/mlx_engine.py).
# mlx-lm is capped at the minor version tested: the backend uses its Qwen3.5 module layout and
# prompt-cache classes, which are internals.
mlx = [
    "mlx>=0.32.2; sys_platform == 'darwin' and platform_machine == 'arm64'",
    "mlx-lm>=0.32,<0.33; sys_platform == 'darwin' and platform_machine == 'arm64'",
]
# `--device cuda`: flash-linear-attention's Triton kernels for the Gated DeltaNet layers. The conv
# layer's kernel, causal-conv1d, compiles against the local CUDA toolkit and is not listed; on a
# machine with nvcc: pip install causal-conv1d --no-build-isolation.
cuda = ["flash-linear-attention>=0.5; sys_platform == 'linux'"]
# `--device mps` and `--device cpu` need nothing beyond the base install. The extras exist so that
# every device installs the same way.
mps = []
cpu = []
dev = [
    "pillow>=10",
    "pytest>=8.0",
    "ruff>=0.6",
    "mypy>=1.15,<3",
    "pre-commit>=3.7",
    "commitizen>=3.28",
]

[project.scripts]
strands-decider = "strands_decider.cli:main"

[tool.setuptools.packages.find]
where = ["src"]

# setuptools-scm drives the version. Release runs override this via
# SETUPTOOLS_SCM_PRETEND_VERSION (see .github/workflows/release-pypi.yml);
# local builds without a tag fall back to 0.0.0, so `pip install -e .` and the
# package-build smoke test still work on any branch.
[tool.setuptools_scm]
fallback_version = "0.0.0"
# Only v<major>.<minor>... tags are package versions. Model release tags
# (model/strands-decider-2B-qwen3.5-v1-2610) are not: without --match,
# `git describe` finds them and the version fails to parse. The dot is required, so a short
# tag such as `v1` or `v1-2610` (a recipe generation, not a package version) is ignored too.
# ponytail: setuptools-scm 10 warns that this key is deprecated (scm.git.describe_command);
# the new key is not read by 8.x, which build-system.requires still admits.
git_describe_command = ["git", "describe", "--dirty", "--tags", "--long", "--match", "v[0-9]*.[0-9]*"]

[tool.ruff]
line-length = 100
target-version = "py310"

[tool.ruff.lint]
# Correctness, import hygiene, and pyupgrade (PEP 585 / PEP 604) modernisation.
select = ["E", "F", "I", "W", "B", "SLF", "RUF", "UP"]
ignore = [
    "B008",    # typer.Option(...) in a default is the idiomatic Typer pattern
    "E501",    # line length is handled by the formatter
]

[tool.ruff.lint.per-file-ignores]
"tests/*" = ["SLF001"]  # tests legitimately assert on internal invariants
"training/tests/*" = ["SLF001"]
"evaluation/*" = ["SLF001"]  # the probes read model internals, as the tests do
"training/data/policy.py" = ["SLF001"]  # attaches _cell/_rule to its own examples for balancing
"research/scripts/*" = ["RUF001"]  # chart labels intentionally use the × multiplication sign

[tool.mypy]
# Matches the strict-ish profile used by strands-agents/harness-sdk. Type-check the
# published surface (`src/strands_decider`) and the training package; the rest of the tree
# (tests, evaluation, research, inference scripts) is deliberately looser. `hatch run
# lint:check` passes `./src ./training --exclude training/tests` to pin that scope.
python_version = "3.10"
warn_return_any = true
warn_unused_configs = true
disallow_untyped_defs = true
disallow_incomplete_defs = true
check_untyped_defs = true
disallow_untyped_decorators = true
no_implicit_optional = true
warn_redundant_casts = true
warn_unused_ignores = true
warn_no_return = true
warn_unreachable = true
follow_untyped_imports = true
ignore_missing_imports = false

# Third-party packages without published type stubs.
[[tool.mypy.overrides]]
module = [
    "torch.*",
    "transformers.*",
    "peft.*",
    "datasets.*",
    "safetensors.*",
    "huggingface_hub.*",
    "yaml",
    "mlx.*",
    "mlx_lm.*",
    "PIL.*",
    "tokenizers",
]
ignore_missing_imports = true

[tool.pytest.ini_options]
testpaths = ["tests", "training/tests"]
# "." so `import training.*` resolves; test_hf_export imports multistep_eval; "tests" and
# "training/tests" so test_variance (in training/tests) can `from test_checkpoint_load` /
# `from test_hf_export` (in tests/) regardless of collection order.
pythonpath = [".", "evaluation", "tests", "training/tests"]
addopts = "-m 'not distributed'"
markers = [
    "distributed: launches torchrun worker processes (minutes); run with -m distributed",
]

[tool.coverage.report]
fail_under = 50

# hatch is used as a dev-workflow runner (envs and scripts) only; the wheel is
# still built with setuptools above. Contributors do not need hatch — the raw
# commands under `## Development` in CONTRIBUTING.md work the same — but if
# they have it, `hatch run test` / `hatch run lint` / `hatch run format` will
# just work.
[tool.hatch.envs.default]
features = ["dev"]

[tool.hatch.envs.default.scripts]
test = "pytest -q {args}"
test-distributed = "pytest -q -m distributed {args}"
format = "ruff format {args:.}"

# Detached lint env: skips the project install (and therefore torch) so `hatch run lint:check`
# stays fast in CI and locally. The typed runtime deps are listed explicitly so mypy can resolve
# their imports; keep this list in sync with [project].dependencies for pydantic/fastapi/typer/rich/uvicorn.
[tool.hatch.envs.lint]
detached = true
dependencies = [
    "ruff>=0.6",
    "mypy>=1.15,<3",
    "pydantic>=2.7",
    "fastapi>=0.115",
    "uvicorn[standard]>=0.30",
    "typer>=0.12",
    "rich>=13.7",
]

[tool.hatch.envs.lint.scripts]
check = ["ruff check {args:.}", "mypy ./src ./training --exclude training/tests"]

# Conventional Commits, matching the PR-title gate in
# .github/workflows/pr-title.yml. Enforced locally by the commit-msg hook in
# .pre-commit-config.yaml.
[tool.commitizen]
name = "cz_conventional_commits"
version = "0.1.0"
tag_format = "v$version"