[build-system]
requires = ["setuptools>=64"]
build-backend = "setuptools.build_meta"

[project]
name = "aileron"
version = "0.1.6"
description = "Flight recorder for AI agents: tamper-evident audit logging, policy enforcement, and incident reports"
readme = "README.md"
license = { text = "Apache-2.0" }
requires-python = ">=3.10"
authors = [{ name = "Kamal Lohchab", email = "kamal.lohchab@yahoo.com" }]
keywords = ["ai-agents", "security", "audit", "mcp", "tamper-evident", "observability"]
classifiers = [
    "Development Status :: 3 - Alpha",
    "Intended Audience :: Developers",
    "Intended Audience :: Information Technology",
    "License :: OSI Approved :: Apache Software License",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.10",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Programming Language :: Python :: 3.13",
    "Topic :: Security",
    "Topic :: System :: Logging",
]
dependencies = [
    "pyyaml",
    "cryptography",
]

[project.urls]
Homepage = "https://github.com/aileron-sh/aileron"
Repository = "https://github.com/aileron-sh/aileron"
Changelog = "https://github.com/aileron-sh/aileron/blob/main/CHANGELOG.md"
Security = "https://github.com/aileron-sh/aileron/blob/main/SECURITY.md"

[project.optional-dependencies]
# ruff is pinned to the exact version CI installs (see .github/workflows/ci.yml)
# so a local run and the CI run always agree. Bump both together.
dev = ["pytest", "ruff==0.16.4"]

[project.scripts]
aileron = "aileron.cli:main"

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

[tool.setuptools.package-data]
aileron = ["py.typed", "rules/examples/*.yml", "rules/examples/*.md"]

# ---------------------------------------------------------------------------
# Lint configuration (ruff). CI runs `ruff check .` and fails on any finding.
# Formatting (`ruff format`) is deliberately not enforced; see CONTRIBUTING.md.
#
# Rule selection, and the reasoning:
#
#   E, F, W   pycodestyle and pyflakes: syntax errors, undefined and unused
#             names, and whitespace errors. The baseline every Python project
#             should have.
#   I         import ordering (isort), so import diffs stay mechanical.
#   UP        pyupgrade: modern syntax that Python 3.10 already supports.
#   B         bugbear: likely bugs such as mutable defaults and unused loop
#             variables.
#   SIM       simplify: readability rewrites that keep behaviour, for example
#             files opened without a context manager (SIM115).
#   RUF       ruff's own rules, including unused `noqa` directives (RUF100)
#             so suppressions cannot silently outlive the finding they hid.
#   S         bandit: this is a security tool, so the security linter runs on
#             it. Findings that are the product's purpose (the proxy spawning
#             the child it mediates) are suppressed one line at a time with a
#             `# noqa` that says why, never with a blanket ignore.
#
# Not selected, on purpose:
#
#   BLE (blind except) and TRY are not selected. The broad `except Exception`
#   blocks in proxy.py, signing.py, mcpserver.py and prefilter.py are
#   deliberate fail-closed catches: on those paths any exception must stop
#   mediation or fail verification rather than propagate. CLAUDE.md explains
#   the posture. A linter that asks for narrower catches there is asking for
#   a less safe product.
#
# `line-length` is 120 because the longest line in the tree measured 118 when
# this configuration landed. The intent is to stop lines growing, not to
# reflow existing code; a lower limit would have forced edits that carry no
# behaviour change and make review harder.
# ---------------------------------------------------------------------------
[tool.ruff]
line-length = 120
target-version = "py310"

[tool.ruff.lint]
select = ["E", "F", "W", "I", "UP", "B", "SIM", "RUF", "S"]
ignore = [
    # try/except/pass on a cleanup path names the exception it swallows;
    # contextlib.suppress would hide that and add a context manager to the
    # proxy's shutdown path for no change in behaviour.
    "SIM105",
    # `list(x)[0]` fails with IndexError on an empty log, `next(iter(x))`
    # fails with StopIteration. Only tests use the former, and the rewrite
    # changes the failure mode for nothing.
    "RUF015",
]
# The prefilter exists because these three codepoints fold to ASCII under
# re.IGNORECASE but not under str.lower() or str.casefold(). The test strings
# and the comments that explain them use the characters on purpose.
allowed-confusables = ["ı", "ſ", "K"]

[tool.ruff.lint.per-file-ignores]
# S101: pytest tests are built out of assert statements.
# S108: tests pass literal /tmp paths as tool-call payloads and as
#       path-traversal probes; nothing is opened there. Real files live in
#       tmp_path fixtures.
# S311: tests use random for fuzz inputs, not for anything secret.
# S603: the live MCP test spawns the proxy under test.
"tests/*" = ["S101", "S108", "S311", "S603"]
# Maintainer scripts, not library code. benchmark.py spawns the child it
# measures and asserts on its own sanity checks; collect_metrics.py talks to
# the GitHub and PyPI APIs by design (the library itself never makes a
# network call, and S310 keeps it honest).
"scripts/*" = ["S101", "S310", "S603"]