wildlint

Static checks distilled from real upstream bugs — the kind off-the-shelf linters miss because they look like ordinary, working code.

Every rule here was born from a concrete bug that was found and fixed in a public project, then generalized to the smallest static check that still catches the class without flooding you with false positives. If a bug could not be turned into a low-noise rule, it is documented as not-shipped rather than added as noise (see Not shipped).

What it catches

Real bugs, phrased the way you'd search them:

• "my argparse flag parses but does nothing" — an option whose dest is never read (WL004)
• x.replace(prefix, "") corrupts values containing the marker twice — meant str.removeprefix/removesuffix (WL001)
• s[-k] raises IndexError on short inputs — deep negative indexing (WL003)
• millify(999999) returns '1000k' not '1M' — rounding rollover in number/byte humanizers (WP001)
• .replace(second=0) crashes on a bare datetime.date — datetime-subclass confusion (WP002)

Install

pip install wildlint

Use

wildlint path/to/code            # scan a file or directory (default: .)
wildlint --select WL001,WL002 src/
wildlint --pedantic src/         # also run opt-in, higher-false-positive rules
wildlint --format json src/      # machine-readable output

When walking a directory, common junk (.venv, __pycache__, build, dist, .git, node_modules, …) is skipped automatically — pass --no-default-exclude to scan everything, or --exclude 'glob/*' to drop more. Explicit file and directory arguments are always scanned as-is. Silence a finding inline with a trailing # noqa (all codes) or # noqa: WL001,WL002 (specific).

Exits non-zero when anything is found or a file could not be analysed (a syntax error, non-UTF-8, or a missing path); the diagnostic goes to stderr and findings stay on stdout, so it drops straight into CI or a pre-commit hook.

pre-commit

# .pre-commit-config.yaml
repos:
  - repo: https://github.com/patchwright/wildlint
    rev: v0.6.0
    hooks:
      - id: wildlint

CI (GitHub Actions)

- run: pip install wildlint
- run: wildlint src/

Configuration

[tool.wildlint] in pyproject.toml sets defaults that CLI flags override:

[tool.wildlint]
pedantic = true          # run opt-in rules by default
select = ["WL001"]       # restrict to these codes
exclude = ["vendor/*"]   # additional path globs to skip

Rules

The default tier is WL001 and WL004 — both have effectively zero false positives. WL002, WL003, and WL005 are opt-in via --pedantic: real bug classes, but they also fire on legitimate code, so the default stays strictly precision.

Each rule is verified against the actual pre-fix source of the project it came from — see the tests, and the rule docstrings in src/wildlint/checkers.py.

Property-test templates

Some bug classes have no stable AST signature — the same wrong behaviour is reached by different code each time, so any static rule broad enough to catch them all also flags mountains of correct code. The archetype is the rounding-rollover bug in number / byte / SI-prefix humanizers (boltons#403, millify#13, numerize#17, si-prefix#17): four distinct implementations of one invariant break (<=-vs-<, a missing carry after rounding, rounding an unrounded boundary). millify(999999) returns '1000k' instead of '1M'.

What they share is a falsifiable property: a humanizer must never emit a mantissa >= base while a larger unit is still available. wildlint ships that check two ways.

Run it directly (dependency-free, in your own test suite or CI):

from wildlint.property_templates import find_rollover
from millify import millify

def test_no_rounding_rollover():
    violations = find_rollover(millify, base=1000)  # 1000=SI, 1024=bytes
    assert not violations, "\n".join(str(v) for v in violations)

find_rollover sweeps the dangerous boundary inputs (values that round up across a unit boundary) and returns the concrete violations. Pass units=[...] (small→large) for an exact check that won't flag legitimate overflow at the largest unit.

The same two-way model covers the date/datetime-subclass confusion bug (deepdiff#602): a function written assuming datetime.datetime that calls .replace(second=0, microsecond=0) (or reads .hour) crashes on a bare datetime.date, because datetime is a subclass of date — so any isinstance(x, date) dispatch admits dates the code cannot handle.

from wildlint.property_templates import find_date_kwargs

def test_does_not_crash_on_date():
    violations = find_date_kwargs(truncate)  # probes with a bare date and time
    assert not violations, "\n".join(str(v) for v in violations)

find_date_kwargs records only TypeError/AttributeError whose message cites a time-only field (hour, minute, second, …); an unrelated crash is a different class and is skipped.

Or render a paste-ready template:

wildlint --template rollover --func millify --import-from millify --base 1000
wildlint --template date-time-kwargs --func truncate --import-from deepdiff
wildlint --template roundtrip --func encodebytes --import-from base62 --inverse decodebytes

Bugs considered but not shipped

Some real bugs do not generalize into a low-false-positive static rule. They are recorded in NON_GENERALIZED in checkers.py so the reasoning is preserved:

• break-vs-continue (mnamer#371) — whether break should be continue is entirely loop-intent dependent.
• sign-doubling (humanize#326) — a numeric-formatting concern, not a syntactic pattern.
• validation-branch-order (validators#463) — specific to one parser's control flow.
• radix-from-ignored-param (shortuuid#115) — requires matching a docstring contract to the implementation.
• rng-from-unordered-set — iterating a set into a random population (directly, or via list(some_set) feeding random.choices weights) is non-deterministic across processes: PYTHONHASHSEED varies per worker, so set iteration order — and item↔weight alignment — changes run to run. The bare form (random.choice({1,2,3})) is rare; the real class (set→list→positional use) is only visible cross-process and is best caught by a reproducibility property test (run twice under differing PYTHONHASHSEED, assert identical output), not a static rule.

Adding a rule

A checker is any object with code, name, tier, and check(tree, path, source=None) -> list[Finding]. Append an instance to CHECKERS in checkers.py and add positive/negative tests mirroring the wild bug. That's the whole extension surface — the suite grows one real bug at a time.

License

MIT.