sloptools

Deterministic code-quality tools for Python codebases. Everything is computed from the source with plain AST analysis (no LLM), so the same code always gives the same results.

Install

pip install sloptools

Requires Python 3.9+.

sloptrace

Is your Python codebase getting messier? sloptrace measures it, and tracks how the numbers change over your git history.

It reports four things:

The approach is adapted from the SlopCodeBench paper (Orlanski et al., 2025).

Usage

sloptrace .                                  # analyse the current checkout
sloptrace /path/to/repo --history            # ...and how it changed over git history
sloptrace . --history --every 20 --max-commits 15   # sample every 20th commit, at most 15
sloptrace . --history --since 2024-01-01 -o out.json
sloptrace . --show-literals                  # also list repeated literals (noisy, off by default)
sloptrace --help                             # all options

Which files are scanned: in a git repo, everything tracked or untracked but not ignored; otherwise the whole directory tree. Tests, docs, build output, virtualenvs and vendored code are skipped by default; change that with --exclude.

Docs

Metric reference: the exact formula and a worked example for each metric.

Compared to scb-check

The paper's authors also publish scb-check. It supports many languages and flags code with lots of pattern rules. sloptrace only handles Python, but gives you evidence (which functions, which copies) and trends over history.

Its erosion number won't exactly match the paper's: the two tools count branches slightly differently, so don't compare absolute values with the paper's published baselines.

slopmerge

Finds functions that have been copied, within one codebase or across several, including copies that have drifted since, and blocks of statements pasted into otherwise different functions. Each set of copies is reported as a group: where every copy is, how similar it is, and what differs between them. Groups are ordered by the lines a merge would save. It's built to feed a Claude skill that decides what to consolidate, but the text summary works on its own too.

slopmerge .                                # duplication within one codebase
slopmerge ../api ../worker ../shared       # ...and between several (all treated as equals)
slopmerge ../api ../worker --cross-root-only   # only code shared between them
slopmerge . --min-similarity 0.8           # only close near-copies (default 0.6)
slopmerge . --units functions,classes      # leave out blocks and constants
slopmerge . --json                         # machine-readable output
slopmerge --help                           # all options

Each group has a tier:

A group is of one kind of unit:

• functions (and methods);
• blocks: runs of statements copied into different functions (extract-a-helper candidates), as exact or renamed copies. A block's differing outside names are the parameters the helper would take;
• constants: the same literal list, tuple, set or dict in several places (column lists, choice tables, config maps). Dicts with the same keys and other values are renamed; near copies list the entries that differ;
• classes: classes declaring the same fields (dataclasses, models, records) or whose methods mostly match (a base class waiting to be extracted). Their matching methods are listed inside the class group.

Code that's duplicated on purpose or by a tool isn't reported: abstract methods and stubs, one-line delegations, property accessors, __init__ methods that only store their arguments, attribute-list __eq__/__repr__/ __hash__, and generated files (protobuf, migrations, files headed "do not edit"). Neither are functions made only of statements found all over the codebase. What was left out is counted in the JSON output.

Each function or block copy also lists what it depends on (uses): the imports, module globals and self attributes it reads, and for a block the enclosing function's variables it needs, i.e. the parameters an extracted helper would take.

See the spec for exactly how each kind of match works.

Development

git clone https://github.com/sjtechdev/sloptools.git
cd sloptools
pip install -e ".[test]"
pytest