pyproject.toml

# Packaging config for PyPI.
#
# A fresh `pip install qikly` is runnable on its own: `qikly --init` creates
# the project layout and a starter task in whatever directory you are standing
# in, and `qikly --scaffold <file.py>` writes a task for code that already
# exists. The bundled tasks, agent definitions, fixtures and default settings
# ship inside the package (see paths.py), so an install and a git clone run
# against the same defaults and produce the same outputs.

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

[project]
name = "qikly"
version = "0.6.1"
description = "Generates a test suite from acceptance criteria, then converges code against it with an agent that never sees those criteria."
readme = "README.md"
requires-python = ">=3.10"
license = "Apache-2.0"
license-files = ["LICENSE"]
authors = [{ name = "Gal Arav" }]
keywords = ["llm", "agent", "testing", "qa", "acceptance-criteria", "code-generation", "mcp", "model-context-protocol"]
classifiers = [
    "Development Status :: 3 - Alpha",
    "Intended Audience :: Developers",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.10",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Topic :: Software Development :: Quality Assurance",
    "Topic :: Software Development :: Testing",
]

# Mirrors requirements.txt. Kept in sync by hand rather than read from that
# file at build time: dynamic dependencies make the built metadata depend on
# files that aren't in the sdist, which is a classic source of "works from a
# clone, broken from PyPI".
# Upper bounds on every one, and the reason is on record rather than
# theoretical: anthropic 1.x removed `temperature` from Messages.create(), so
# a provider that had worked stopped working on a major bump nobody here made.
# A floor alone lets that happen silently to a user on a fresh install.
#
# These are deliberately wide. The bound exists to stop a major version
# arriving unannounced, not to pin a patch release, and raising one after
# testing is a two-line change.
dependencies = [
    "PyYAML>=6.0,<7",
    "google-genai>=2.0,<3",
    "pytest>=8.0,<10",
    "rich>=13.0,<16",
]

# Not installed by default: pick the one you use. Only Gemini is a hard
# dependency, and tests/test_provider_signatures.py checks whichever of these
# happens to be present against what the code sends it.
[project.optional-dependencies]
openai = ["openai>=1.0,<4"]
anthropic = ["anthropic>=0.30,<2"]
all-providers = ["openai>=1.0,<4", "anthropic>=0.30,<2"]
# The MCP server. Optional because it is a way in, not the tool: everything it
# exposes works without it, and every test in this repository runs without it.
mcp = ["mcp>=2.0,<3"]

[project.urls]
Homepage = "https://test.qikly.com/"
Documentation = "https://test.qikly.com/"
Repository = "https://github.com/gal-a/qikly"
Issues = "https://github.com/gal-a/qikly/issues"
Changelog = "https://github.com/gal-a/qikly/blob/main/CHANGELOG.md"

[project.scripts]
qikly = "qikly.cli:main"
qikly-all = "qikly.orchestrator.run_all:main"
qikly-report = "qikly.orchestrator.reports.report:main"
qikly-metrics = "qikly.orchestrator.reports.metrics_report:main"
qikly-diff-criteria = "qikly.orchestrator.tuning.diff_acceptance_criteria:main"
qikly-mcp = "qikly.mcp_server:main"

[tool.setuptools]
# src layout: the package lives in src/, not at the repo root. That keeps the
# project's own inputs/, inputs_private/, and outputs/ from sitting next to
# importable code, and it makes it impossible to import qikly from the
# source tree by accident when you meant to exercise the installed copy --
# which is exactly the mistake that hides a broken wheel until after upload.
package-dir = { "" = "src" }

# Explicit, not auto-discovered: auto-discovery would sweep in `outputs` and
# any `inputs_private` the developer has locally as if they were packages.
packages = [
    "qikly",
    "qikly.agent_api",
    "qikly.agent_api.code_loader",
    "qikly.agent_api.prompts",
    "qikly.agent_api.providers",
    "qikly.agent_tools",
    "qikly.orchestrator",
    "qikly.orchestrator.reports",
    "qikly.orchestrator.tuning",
]
# No py-modules. The root run.py is a clone convenience only -- installing a
# top-level module named `run` would reintroduce exactly the collision that
# nesting everything under qikly/ removes.

[tool.setuptools.package-data]
# The bundled defaults ship in the wheel. This is what makes a fresh
# `pip install` immediately runnable: the example tasks, agent definitions,
# sample data, and default settings.yaml are all present,
# so a new user's first run produces the same outputs as a clone's.
# Their own specs go in inputs_private/ in their project directory, which is
# never packaged -- see qikly/paths.py.
"qikly" = [
    "inputs_public/config/*.yaml",
    "inputs_public/config/tasks/*.yaml",
    "inputs_public/reference/*/*.py",
    "inputs_public/agent_defs/*.md",
    # The agent Skill, so `qikly --install-skill` works from a wheel and not
    # only from a clone. A Skill is a folder an agent reads, and its two
    # bundled references are complete copies of the docs pages, asserted
    # identical by tests/test_skill_bundle.py, so it works with no network.
    "skills/*/*.md",
    "skills/*/references/*.md",
    "inputs_public/data/**/*",
    # The worked scaffold example, laid out the way a real project is so a
    # reader sees the right shape. Not a bundled task: it lives under
    # examples/ so it stays out of the task list and out of --validate.
    "inputs_public/examples/**/*",
    # The MCP server's icon. Shipped rather than read out of docs/, which is
    # not in the wheel, and embedded as a data URI at runtime so no host has
    # to resolve a path on the reader's filesystem.
    "assets/*.svg",
]

[tool.pytest.ini_options]
# tests/ only. Without this, a bare `pytest` also collects outputs/tests/,
# which holds generated suites from previous runs: they import task modules
# that may not exist, and their failures say nothing about this package.
testpaths = ["tests"]
# No -q here. A project addopts of "-q" combines with a "-q" the user types
# into -qq, and -qq suppresses the "N passed" summary line entirely: the run
# then reports a bare "100%" and says nothing about how many tests ran or
# whether any failed. That is not hypothetical. The same collision inside this
# tool's own pytest calls silently zeroed an entire experiment's measurement,
# which is why those calls now pass "-o addopts=". Anyone who wants quiet
# output can still type -q and get the summary with it.