Check how long your functions and classes are.

A long function is the cheapest smell to detect and the most reliable one to act on. dlen tells you which ones crossed the line, and where.

$ dlen src/
src/importer.py:42:1: DL001 function 'process_batch' is 61 lines (max 40)
src/importer.py:88:5: DL001 function 'validate' is 28 lines (warn 25)
src/models.py:7:1: DL002 class 'LegacyRecord' is 812 lines (max 500)

It counts the code you have to read: not the docstring, not the blank lines.

uv tool install dlen     # or: pipx install dlen, pip install dlen

Python 3.10 or newer. No dependencies.

dlen src/ tests/                 # files, directories, or both
dlen . --max-function 30         # your project, your thresholds
dlen . --output-format=json      # for other tools

Decorators count — they are part of what you read before you understand the function.

Docstrings and blank lines do not. Counting them meant punishing you for documenting: a well-written docstring pushed a short function over the limit. rich.inspect is 54 lines on screen and 7 statements of code; flagging it was simply wrong.

Comments do count. They are text you read to understand the code, not text that describes the interface.

They were measured over 28,390 functions in the Python standard library, numpy, Pillow, rich, pytest, httpx, mypy and coverage, counted the way above:

The median function out there is 5 lines, and that holds across all eight codebases. --max-function sits where it flags about 5% — a list you can work through on a Tuesday — and --warn-function around 10%, close enough to notice before it becomes a problem.

The class limit is the one number from 2017 the data supports: 500 flags 2.2% of classes, and a class that long really is a god object. It stays.

If you want the stricter Clean Code reading, it is one flag away:

dlen src/ --warn-function 12 --max-function 20

Versions before 0.2.0 shipped 12 and 20, and 0.2.0 raised them to 30 and 50. 0.3.0 stopped counting docstrings and blank lines, which made every function measure shorter, so the thresholds came down with it.

Warnings never fail the build. That is the point of having two thresholds: one to nudge, one to stop.

--output-format=github puts each finding on the diff of the pull request instead of burying it in a log:

- run: uvx dlen src --output-format=github

The core returns data and prints nothing, so you can build your own reporting on it:

from pathlib import Path
from dlen import Limits, check_paths

findings = check_paths([Path("src")], Limits(max_function=30))
for finding in findings:
    print(finding.code, finding.path, finding.line, finding.message)

Every call returns a fresh list. Two runs never see each other's findings.

dlen parses your code with Python's own ast module, so it agrees with Python about what a function is. It sees async def, decorators, nested functions and methods, and it is not fooled by the word def inside a string or a comment.

Versions before 0.1.0 matched text with regular expressions and had been broken on Python 3 since 2017. 0.1.0 is a rewrite: the counts are now correct rather than compatible, and the output carries file, line and column.

MIT.