LaNorme

LaNorme makes a codebase's standard executable. It automates the mechanical side of code review: on every commit it checks cyclomatic complexity, file and function size, duplication, stale doc references, architectural boundaries, and naming conventions, and fails the build when they drift.

The built-in normes cover the common ground. A plugin interface lets a team encode its own, so the standard you agree on is the standard the build keeps. The same gate runs in CI and inside an AI agent's loop, so people and agents write to one bar and the codebase stays clean as it grows.

Standard library only. No runtime dependencies. Python 3.13+.

Full documentation: lanorme.github.io/lanorme. Tutorials, how-to guides, the complete rule and configuration reference, and agent-friendly Markdown: every page is also served raw at its .md URL, with an llms.txt index for agents.

Why LaNorme

• Keep a growing codebase clean. Complexity, size, duplication, and stale docs are caught the moment they appear, while the fix is still small.
• Enforce architecture and conventions. Layering, ports-and-adapters wiring, and your own domain vocabulary become checks any contributor can run.
• Gate AI-generated code. Hand an agent the same normes your team codes to: it gets concrete, mechanical feedback on what good looks like here, and non-compliant output fails before it reaches a human.
• Spend review on judgement. Reviewers focus on design and correctness because the mechanical checks are already green.

Install

uv tool install lanorme       # or: pipx install lanorme, pip install lanorme
uvx lanorme check .           # or run once without installing

Releases are tagged vX.Y.Z; see the releases page for notes and the documentation for other install methods.

Quickstart

lanorme check [PATHS...]               # run every enabled check (default path: .)
lanorme check . --check secrets        # run one check by name, rule code, or category
lanorme check . --select TYPE,AUTHN    # only these rule codes or categories
lanorme check . --output-format ndjson # one finding per line, for jq / grep
lanorme rules                          # list every registered rule

Exit code is 1 when any check fails, 0 when the tree is clean. By default a run reports only the checks that found something, then a summary line:

$ lanorme check src/
[FAIL] secrets
  VIOLATION: app.py:8 — Hardcoded credential value bound to 'API_KEY'
    Rule: SECRETPY-001: No hardcoded secrets in source code
    Fix: Read the value from an environment variable, secrets manager, or settings module
--- secrets: 1 violations, 0 warnings ---

Summary: 24 checks — 23 passed, 0 warnings, 1 failed.

The full command and flag reference and output formats are in the CLI reference. Inline suppression, # noqa and the ruff-safe # lanorme: ignore[CODE], is in Configure which checks run.

What it checks

lanorme rules prints the live list. The rule reference documents every rule: what it catches and what it does not, its config, and its measured precision and recall on the bundled corpora.

On by default, on any project, no config needed:

Off until you turn them on (layered or hexagonal architecture, domain vocabulary, house styles, Markdown docs structure, and experimental precision-first detectors): LAYER, PORT, TERM, KWARG, NAMING-001/002, AAA, CMT-005, SIMILAR, ATTR, PROSE, DOCS, PATH, STALE. The rule reference documents each.

Configuration

LaNorme walks up from the target path looking for config: a dedicated lanorme.toml, otherwise a [tool.lanorme] table in pyproject.toml. Command line flags win over both.

[tool.lanorme]
extends = ["strict", "hexagonal"]   # adopt bundled profiles; local keys win
select = ["ALL"]
ignore = ["NAMING-003"]
promote = ["TYPE-004"]              # advisory warnings become build-failing errors
exclude = ["postman/**", "vendor/*"]

That is the surface. The docs cover the rest without repeating it here:

• Configuration reference: every key, its type and default, plus a machine-readable JSON schema.
• Profiles (extends): the strict, hexagonal, clean and layered bundles.
• Per-directory config: drop a nested lanorme.toml to tighten one subtree while the rest stays lenient.

Adopting on an existing codebase

A mature codebase has findings on day one. A baseline records the debt you already have so only new findings report; the whole adoption is two commands:

lanorme baseline write    # records current findings to lanorme-baseline.json

Add baseline = "lanorme-baseline.json" under [tool.lanorme] and commit the file like a lockfile. From then on every check runs at full strictness, but only what you add reports. The full walkthrough is the adopt-on-an-existing-codebase tutorial.

Writing a check

A check is any object with name, description, rules, and a run method; drop it in lanorme/checks/, ship it under the lanorme.checks entry-point group, or point at it with [tool.lanorme] plugins = [...]. The write-a-check guide and CONTRIBUTING.md cover the setup, the gates, and the conventions for a new rule.

Versioning

The public surface is the rule codes you put in select / ignore / per-file-ignores and the config keys under [tool.lanorme]. The question that decides a bump is whether a green codebase could go red on upgrade: a patch keeps every result unchanged, a minor can newly fail a previously-passing codebase (every pre-1.0 breaking change lands here), and a major is the stability commitment. Every change is listed in CHANGELOG.md.

A rule's human-readable description is not part of that surface and may be reworded in a minor release. A file-level finding anchors on its description, so rewording one detaches the matching entries in a committed baseline and those findings report again until you run lanorme baseline write once. The run tells you when this has happened, naming the file and rule rather than letting old debt look new, and the changelog entry says so for the release that causes it.

Licence

MIT. See LICENSE.