[project]
# Import package and CLI are `distil`; the PyPI distribution is `distil-llm`
# (the bare `distil` name is already taken on PyPI).
name = "distil-llm"
version = "1.53.0rc1"
description = "Compression with a quality contract — cache-aware, causally-pruned context compression for agentic runtimes, gated by a statistical non-inferiority test."
readme = "README.md"
# Python floor is 3.9 — the version macOS still ships as the system `python3`. The
# package is stdlib-only and uses `from __future__ import annotations`, so it imports
# and the `distil bench` gate passes on 3.9/3.10/3.11/3.12/3.13 (verified in CI). A
# higher floor here is what made `pipx install distil-llm` fail with the misleading
# "Could not find a version ... (from versions: none)" on stock macOS Python.
requires-python = ">=3.9"
license = "Apache-2.0"
license-files = ["LICENSE"]
authors = [{ name = "shakes", email = "chandu1221@gmail.com" }]
keywords = ["llm", "context-compression", "prompt-caching", "agentic-runtime", "non-inferiority", "tokens", "cost"]
classifiers = [
  "Development Status :: 5 - Production/Stable",
  "Intended Audience :: Developers",
  "Topic :: Scientific/Engineering :: Artificial Intelligence",
  "Programming Language :: Python :: 3.9",
  "Programming Language :: Python :: 3.10",
  "Programming Language :: Python :: 3.11",
  "Programming Language :: Python :: 3.12",
  "Programming Language :: Python :: 3.13",
  "Operating System :: OS Independent",
  "Typing :: Typed",
]
dependencies = []  # core is stdlib-only by design: it must run anywhere, day one.

[project.urls]
Homepage = "https://github.com/dshakes/distil"
Repository = "https://github.com/dshakes/distil"
Issues = "https://github.com/dshakes/distil/issues"
Changelog = "https://github.com/dshakes/distil/blob/main/CHANGELOG.md"

[project.optional-dependencies]
# Production hardening swaps these in; the core does not require them.
#   - live: billing-grade tokenizer (count_tokens) + live AgentRunner, via the Anthropic SDK
dev = ["pytest>=8", "opentelemetry-api>=1.20", "opentelemetry-sdk>=1.20"]
live = ["anthropic>=0.40", "openai>=1.60"]
onnx = ["onnxruntime>=1.17", "transformers>=4.40"]   # transformer keep-model inference
train = ["torch>=2.2", "transformers>=4.40", "onnx>=1.16", "onnxscript>=0.1", "sentencepiece>=0.2"]  # training + ONNX export
async = ["aiohttp>=3.9"]                              # async high-concurrency proxy (distil.aproxy)
otel = ["opentelemetry-api>=1.20"]                    # GenAI semantic-convention spans (distil.otel)
# OIDC RS256 verification for the gateway. HS256 needs nothing (stdlib hmac);
# this is only for asymmetric IdP signing keys. Without it an RS256 token is
# REFUSED, never verified-by-skipping — see distil/authz.py.
oidc = ["cryptography>=42.0"]
# Image downscaling (distil/compress/vision_scale.py). Resampling needs a real
# codec, which the stdlib does not have. Without this extra the transform is
# INERT — available() is False and images pass through untouched — so distil's
# zero-dependency guarantee is unchanged for anyone who does not opt in.
image = ["Pillow>=10.0"]
# native = the Rust hot-path core (rust/distil-core); build with `maturin develop --release`

[project.scripts]
distil = "distil.cli:main"
# Unambiguous MCP launch target for registry/client configs: `uvx --from
# distil-llm distil-mcp` — the console script is otherwise `distil` while the
# distribution is `distil-llm`, and MCP server.json launch specs don't cover
# the `script subcommand` form cleanly.
distil-mcp = "distil.mcp_server:serve"
# `uvx <pkg>` runs the executable *named after the package*, so without this the
# official MCP registry's own launch spec (`uvx distil-llm ...`) fails with
# "An executable named `distil-llm` is not provided by package `distil-llm`".
# serve() ignores argv, so the already-published `uvx distil-llm mcp` form works
# too — this repairs the live registry entry without needing it republished.
distil-llm = "distil.mcp_server:serve"

[build-system]
# PINNED, like ruff and mypy in CI, and for the same reason: an unpinned build
# backend drifts and breaks the release. hatchling 1.32 emits
# `Metadata-Version: 2.5`, which the publish action's bundled twine rejects with
# "not a valid metadata version" — so v1.43.0rc3 built green and published
# nothing. Worse, it was invisible locally: this machine had an older hatchling
# cached, so `uv build` produced an acceptable 2.4 wheel while CI produced an
# unpublishable one. Local and CI must build the SAME artifact.
# Lifting this pin requires the publish action to accept 2.5 first — see the
# metadata assertion in tests/test_packaging_smoke.py.
requires = ["hatchling<1.32"]
build-backend = "hatchling.build"

# Keep local cruft (alt venvs, caches, result dumps) out of the sdist — an
# untracked .venv-*/ with absolute symlinks otherwise breaks `uv build`.
[tool.hatch.build.targets.sdist]
exclude = [".venv*", ".e7_cache", "logs", "benchmarks/results*", "docs/paper/results"]

[tool.hatch.build.targets.wheel]
packages = ["distil"]

# Ship the trajectory corpus inside the wheel so `pip install distil && distil bench`
# is self-contained.
[tool.hatch.build.targets.wheel.force-include]
"corpus" = "distil/_corpus"

[tool.ruff]
target-version = "py311"
line-length = 100

[tool.ruff.lint.per-file-ignores]
"tests/test_aproxy.py" = ["E402"]  # imports follow pytest.importorskip("aiohttp")

[tool.pytest.ini_options]
# All tests live under tests/; keep collection out of the SWE-bench E7 harness's
# transient logs/ and image/worktree caches (which contain files named test_*.py).
testpaths = ["tests"]
norecursedirs = ["logs", ".e7_cache", ".venv*", "docs", "corpus", "benchmarks"]

[tool.coverage.run]
branch = false
# Excluded from the coverage metric because they cannot be exercised by an
# offline unit test — they require a GPU/torch training run or live, credentialed
# model calls (the proof-harness runners). They ARE integration-tested manually.
omit = [
    "distil/codec/train_transformer.py",       # torch training loop (optional [transformer] extra)
    "distil/replay/anthropic_runner.py",       # live Anthropic SDK + credentials
    "distil/replay/openai_runner.py",          # live OpenAI SDK + credentials
    "distil/replay/claude_cli_runner.py",      # spawns the real `claude` CLI subprocess
    "distil/replay/expand_runner.py",          # live model round-trip
]

[tool.coverage.report]
# Compare fail_under at 2-decimal precision. coverage.py defaults precision to 0,
# which ALSO rounds the number used for the --cov-fail-under decision (not just
# the printed report): 94.90% rounds to 95 and silently passes a 95 floor. Pin it
# so the gate enforces the real figure. (Issue #32.)
precision = 2
exclude_also = [
    "if TYPE_CHECKING:",
    "raise NotImplementedError",
    "if __name__ == .__main__.:",
    "@(abc\\.)?abstractmethod",
    # Blocking server-launch / subprocess-spawn entry points. The request-
    # handling LOGIC they wire up (build_handler / make_app) is unit-tested
    # directly; these functions themselves run a server forever or Popen the
    # agent, so they're exercised by integration, not unit tests.
    "def serve\\(",
    "def wrap_run\\(",
    "def serve_gateway\\(",
    # Fail-open defensive handlers, explicitly marked `# noqa: BLE001`: they exist
    # ONLY to guarantee "telemetry/compression must never break a request" and do
    # nothing but log at debug and continue. There is no behaviour to assert, and a
    # unit test cannot reliably force the internal exception that triggers them.
    "except .*# noqa: BLE001",
]