pyproject.toml

[project]
name = "contemplative-agent"
version = "2.12.0"
description = "A Python CLI agent that self-updates skills, rules, ethics, and identity from experience"
requires-python = ">=3.10"
license = {text = "MIT"}
authors = [
    {name = "Tatsuya Shimomoto"},
]
keywords = [
    "ai-agent", "autonomous-agent", "self-improving", "local-llm",
    "ollama", "memory-distillation", "contemplative-ai",
    "cli-agent", "ai-safety", "secure-by-design",
]
classifiers = [
    "Development Status :: 4 - Beta",
    "Intended Audience :: Science/Research",
    "License :: OSI Approved :: MIT License",
    "Programming Language :: Python :: 3",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
]
dependencies = ["requests>=2.33.0", "numpy>=1.24.0"]

[project.urls]
Homepage = "https://github.com/shimo4228/contemplative-agent"
Repository = "https://github.com/shimo4228/contemplative-agent"
"Bug Tracker" = "https://github.com/shimo4228/contemplative-agent/issues"

[project.scripts]
contemplative-agent = "contemplative_agent.cli:main"

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

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

[tool.pytest.ini_options]
testpaths = ["tests"]
# The eval group's deepeval plugin prints skipped nodeids as rich markup and dies with
# MarkupError on Linux (CI run 36055176384, 2026-09-24); evals/ never uses the plugin.
addopts = "-p no:deepeval"
markers = [
    "unit: Unit tests",
    "integration: Integration tests",
    # Spawns the real `claude` binary to read a contract no stub can answer for
    # (which flags exist; which `--tools` names the CLI actually honours). Runs
    # by default so an operator's full Verify sees the drift; deselected inside
    # weekly-pipeline.sh's fix loop, where an unrelated CLI hiccup would abort
    # `pytest -x` and be attributed to the fix under test. Every test that
    # spawns the binary must carry it — the exclusion is by class, not by name.
    "live_cli: Spawns the real claude CLI (external binary, no model call)",
]

[tool.pyright]
include = ["src", "tests", "evals"]
venvPath = "."
venv = ".venv"
# Argparse handler signatures take (args, parser) uniformly via dispatcher;
# many handlers do not use parser, so unused-variable and unreachable reports
# are disabled here (the `_` prefix marks intent at the call sites).
reportUnusedVariable = "none"
reportUnreachable = "none"

[dependency-groups]
# Dev tooling lives here (uv auto-syncs the `dev` group on `uv run` / `uv sync`),
# so `uv run pytest` / `uv run ruff` / `uv run pyright` work without extra flags.
dev = [
    # Floor is a security floor, not a feature floor: 9.0.2 and below carry
    # PYSEC-2026-1845 (pip-audit, 2026-07-31).
    "pytest>=9.0.3",
    "pytest-cov>=4.0",
    "hypothesis>=6.100",
    "responses>=0.23.0",
    "types-requests>=2.31",
    # Floor tracks the version the PostToolUse autofix hook resolves, so
    # `uv run ruff` and the hook agree on more than just the rule set.
    "ruff>=0.16",
    "pyright>=1.1.380",
    "import-linter>=2.0",
    # 機械ゲート (.claude/verify.sh) が使う。版の正本をここに集約し、ローカルと CI で
    # 同じものが走るようにする。選定根拠と再調査トリガーは .claude/verify.md
    "bandit>=1.9.4",
    "pip-audit>=2.9",
    # 週次 dead-code intake (scripts/dead_code_scan.py) が使う。ruff が構造的に
    # 検出できない unused function/class/method/attribute を担う（選定 2026-08-07、
    # search-first: MIT / stdlib+toml のみ / 2026-03 v2.16）。検出のみ — 削除は
    # 土曜ゲートの人間承認（T-DEADCODE-INTAKE）。
    "vulture>=2.16",
    # scripts/instrument_census.py の週次集計が使う（ADR-0110）。session × category の
    # 行列・分バケット・間隔 diff・markdown 表は手書き ~150 行が各 1 行になる。
    # wheel には入れない — scripts/ は配布物の外で、依存の重さは wheel の中でだけ
    # 却下理由になる（ADR-0109）。統計は既存の numpy 3 行で足り、scipy は入れない。
    "pandas>=2.2",
    "tabulate>=0.9",      # DataFrame.to_markdown() の実装
    "pandas-stubs>=2.2",  # 型（pyright の include に scripts は無いが、編集時に効く）
]
# LLM behavioral eval 層 (evals/) 専用。uv の既定同期に含めず、eval 実行と
# verify.sh の type ゲートだけが --group eval で取り込む (ADR-0089)。
# pin は pyproject が正本 (uv.lock は gitignore)。
eval = [
    "deepeval==4.1.5",
]

[tool.uv]
# Security floors for packages we do not depend on directly (they arrive via
# requests / pytest / pip-audit). uv.lock is gitignored here, so a lock bump
# alone would not survive a clone — these constraints are the durable record
# of the 2026-07-31 pip-audit drain. Drop an entry once the ecosystem floor
# has moved past it on its own.
constraint-dependencies = [
    "urllib3>=2.7.0",    # PYSEC-2026-141, PYSEC-2026-142
    "idna>=3.15",        # PYSEC-2026-215
    "pygments>=2.20.0",  # PYSEC-2026-2987
]

[tool.vulture]
# 週次 dead-code intake (scripts/dead_code_scan.py) と手動 `uv run vulture` の共通正本
# (T-DEADCODE-INTAKE)。tests/ evals/ は参照解決のために scan へ含める — テストだけが
# 使う src コードを偽陽性にしないため。報告の絞り込み (src/ scripts/ のみ) は
# dead_code_scan.py 側の決定論コードが行う。偽陽性の免除は .vulture_whitelist.py
# (追記・削除は土曜ゲートの人間 commit)。検出のみ — 自動削除は構造的に存在しない。
paths = ["src", "scripts", "tests", "evals", ".vulture_whitelist.py"]

[tool.importlinter]
root_packages = ["contemplative_agent"]

# ADR-0001: one-way dependency cli -> adapters -> core.
# On violation, move the shared code down into core/ (or inject it via
# constructor / SessionContext) instead of importing upward.
[[tool.importlinter.contracts]]
name = "Layered architecture: cli -> adapters -> core (ADR-0001)"
type = "layers"
layers = [
    "contemplative_agent.cli",
    "contemplative_agent.adapters",
    "contemplative_agent.core",
]

# ADR-0015: one external adapter per agent — adapters must stay separable.
# On violation, move the shared logic down into core/ rather than importing
# a sibling adapter.
[[tool.importlinter.contracts]]
name = "Adapter independence (ADR-0015): moltbook / meditation / dialogue"
type = "independence"
modules = [
    "contemplative_agent.adapters.moltbook",
    "contemplative_agent.adapters.meditation",
    "contemplative_agent.adapters.dialogue",
]

# ADR-0088: the conformance kit ships in the wheel (siblings import it) but
# is not production code. Two forbidden contracts rather than a fourth
# `layers` entry: layers would state the direction correctly and ALSO permit
# testing -> adapters, which would drag the Moltbook HTTP client into a
# sibling's test dependencies. What is wanted is a narrow ban on each side.
[[tool.importlinter.contracts]]
name = "Testing kit is a leaf (ADR-0088): production layers never import it"
type = "forbidden"
source_modules = [
    "contemplative_agent.cli",
    "contemplative_agent.adapters",
    "contemplative_agent.core",
]
forbidden_modules = ["contemplative_agent.testing"]

[[tool.importlinter.contracts]]
name = "Testing kit depends on core only (ADR-0088)"
type = "forbidden"
source_modules = ["contemplative_agent.testing"]
forbidden_modules = [
    "contemplative_agent.cli",
    "contemplative_agent.adapters",
]

[tool.ruff]
target-version = "py310"
line-length = 100
# docs/evidence is frozen verbatim records — data, not owned code (same standard
# as the markdown gate's exclusion list in .claude/verify.md). Excluding it here
# covers lint AND format, in both verify.sh modes, and ends the per-rule
# exemption creep (T20 -> C901 -> ... one per-file-ignores line per new rule).
# force-exclude extends this to explicitly-passed paths (the PostToolUse
# autofix hook invokes ruff on the edited file directly).
extend-exclude = ["docs/evidence"]
force-exclude = true

[tool.ruff.lint]
# Pin the rule set explicitly rather than inheriting ruff's defaults. The
# PostToolUse autofix hook resolves ruff from PATH/uvx while this repo locks
# its own version, and the two drifted: a wider default set fired UP/SIM/RUF
# on the files that happened to be edited, and `ruff check src/ tests/` at
# the locked version reported those same files clean — so the tree silently
# split into modernized and legacy halves. An explicit `select` makes the
# answer identical under both binaries (verified: 674 findings either way).
#
# E4/E7/E9/F are ruff's own defaults, restated here so nothing depends on
# what "default" means in a given release. Beyond them: B (bugbear —
# silent-truncation zip, mutable defaults), I (import sorting), T20 (no
# print in library code — use logging; rules/python/hooks.md), UP (pyupgrade
# — PEP 585/604 typing; the floor is py310 so `list[str]` / `X | None` are
# native at runtime). C901 (mccabe) is the complexity budget — a per-function
# ceiling so that "one more branch" has to be argued for rather than merged by
# default; its threshold and exemption ledger are configured below.
# ANN001-003/201 require explicit types at function boundaries and ANN401 bans
# bare `Any` there (LLM-first axis 4, 2026-08-31): a typed boundary is a
# machine-checkable spec — the frozen edge for regeneration and cross-session
# edits. Production code had zero missing annotations at adoption, and the 24
# bare-Any sites were drained the same day with no exemption and no inline
# noqa. ANN202/204-206 (private/dunder return types) stay unselected — the
# explicit-pin policy adds rules when argued for, not by family.
select = [
    "E4", "E7", "E9", "F", "B", "I", "T20", "UP", "C901",
    "ANN001", "ANN002", "ANN003", "ANN201", "ANN401",
]

[tool.ruff.lint.mccabe]
# Budget rule: drain, do not raise — any change to this number needs a dated
# reason in .claude/verify.md (harness ADR-0056). The exemption ledger in
# per-file-ignores is empty and stays empty: a function over the budget is a
# design question, not a config entry. When one cannot be drained the answer
# is to stop lowering the threshold, not to exempt the function.
# Distribution as-of 2026-08-31 (src/tests/scripts/evals): p50=1 / p90=3 /
# p95=5 / p99=9 / max=10. Introduced at 15 on 2026-08-28 (then p99=10 / max=35,
# 13 functions drained the same day) and walked down 15 -> 13 -> 11 -> 10 on
# 2026-08-31, draining 24 functions and exempting none. The budget now sits at
# p99: 26 functions are at 10, so the next branch added to any of them has to be
# argued for rather than merged by default.
# "10 was rejected" in the 2026-08-28 note meant rejected as an INTRODUCTION
# threshold, where it would have put 24 files on the exemption list and
# blinded the most-edited modules. Reached by draining instead, it exempts
# nothing — which is the road that was taken.
max-complexity = 10

[tool.ruff.lint.isort]
# Keep explicit `X as X` re-export blocks (core/llm/__init__.py) compact
# instead of one import statement per name.
combine-as-imports = true
# Declare the first-party roots instead of letting ruff infer them from what
# happens to exist on disk. `.claude/verify.sh --staged` checks the *index*
# by materializing only the staged files into a tmpdir, so a commit that
# stages tests without src/ would leave `contemplative_agent` unresolvable
# there — inferred as third-party, and every test file then fails I001 on a
# section split that full mode considers correct (2026-07-31, 14 false
# positives). Declaring it makes staged and full mode agree by construction.
known-first-party = ["contemplative_agent", "tests", "evals"]

[tool.ruff.lint.per-file-ignores]
# Console-UI seams where print IS the interface, not stray logging.
"src/contemplative_agent/cli/**" = ["T20"]                      # CLI user output
"src/contemplative_agent/adapters/moltbook/agent.py" = ["T20"]  # ADR-0012 approval-gate TTY display
"src/contemplative_agent/adapters/dialogue/peer.py" = ["T20"]   # stderr turn progress
"src/contemplative_agent/testing/__main__.py" = ["T20"]         # ADR-0088 conformance CLI
# tests/ and the ANN family: test functions are not consumed boundaries (nothing
# regenerates against their signatures), and the corpus carries 5,810 missing
# annotations — cost with no LLM-first payoff. Permanent, not a drain ledger.
"tests/**" = ["T20", "ANN001", "ANN002", "ANN003", "ANN201", "ANN401"]
"scripts/**" = ["T20"]
"evals/**" = ["T20"]    # probe/CLI 系 (benchmark_distill と同じ扱い) — print が UI
# docs/evidence/** の免除行は 2026-08-31 に撤去 — [tool.ruff] の extend-exclude が
# ディレクトリごと検査対象から外した (凍結記録はデータであってコードではない)