agent-slop-lint 0.5.0

Agentic code quality linter — catches slop before it becomes architectural rot

slop

A code quality linter for codebases where AI agents are writing most of the diffs.

Static analysis tools got their defaults from codebases where a productive human wrote maybe 100 lines on a busy day and another human reviewed every one. An agent can drop that much into a single file before emitting its first status message, and the structural damage (deep coupling, WMC-heavy classes, files that grow 500 LOC in a week) lands inside one session rather than accumulating over quarters. The usual review cadence does not catch it. slop is calibrated for that pace.

The metrics themselves are not new. Cyclomatic complexity (McCabe 1976), the CK suite for classes (Chidamber and Kemerer 1994), package distance from the Main Sequence (Martin 1994), churn-weighted hotspots (Tornhill 2015): all well-cited, all mostly ignored in day-to-day workflows because they were tuned for human timescales. slop wraps them behind one CLI with thresholds that assume a different pace of change.

Example

$ slop lint

slop 0.5.0 — scanning .

complexity
  cyclomatic
    ✗ src/pipeline/ingest.py:44 process_batch — CCX 18 exceeds 10
    ✗ src/pipeline/ingest.py:112 _normalize_rows — CCX 14 exceeds 10
    ✗ src/store/frame.py:204 append_partition — CCX 11 exceeds 10

  cognitive
    ✗ src/pipeline/ingest.py:44 process_batch — CogC 26 exceeds 15
    ✗ src/store/frame.py:204 append_partition — CogC 21 exceeds 15

  6 violations, 142 checked

hotspots (14 days ago)
  ✗ src/lifecycle/tasks/write.py — CCX=41, growth +556 LOC
  ✗ src/pipeline/transformation.py — CCX=45, growth +367 LOC
  2 violations

packages
  ⚠ src/config — Zone of Pain (I=0.12, A=0.00)
  ⚠ src/core/distributed/dag — Zone of Pain (I=0.18, A=0.05)
  ⚠ src/core/transform — Zone of Pain (I=0.09, A=0.03)
  3 advisories

deps
  ✓ clean, 55 checked

class
  coupling
    ✗ tests/v2/test_pdrf_types.py:12 PdrfTestSuite — CBO 11 exceeds 8

  1 violation, 8 checked

orphans (disabled)
  ℹ skipped (enable in .slop.toml)

────────────────────────────────────────
9 violations | 3 advisories | 8 rules checked | FAIL

Exit code is 0 when clean, 1 on violations, 2 on configuration or runtime error. Works in CI, pre-commit hooks, and interactively.

Install

pip install agent-slop-lint

slop shells out to rg, fd, and git. Install those via your system package manager (apt install ripgrep fd-find git, brew install ripgrep fd git, or equivalent) and run slop doctor to verify.

Full per-platform install steps, CI recipes, pre-commit wiring, and the agent skill are in the setup guide.

Rules

Rule	Default	Measures	Source
complexity.cyclomatic	CCX > 10	Per-function path count	McCabe 1976
complexity.cognitive	CogC > 15	Per-function reading difficulty	Campbell 2018
complexity.weighted	WMC > 40	Per-class aggregate method complexity	Chidamber and Kemerer 1994
halstead.volume	V > 1500	Per-function information content (Length × log2 Vocabulary)	Halstead 1977
halstead.difficulty	D > 30	Per-function operator/operand density	Halstead 1977
npath	NPath > 400	Per-function acyclic execution path count	Nejmeh 1988
hotspots	14-day window	Files that are complex AND growing fast	Tornhill 2015
packages	D' > 0.7	Package design distance from the Main Sequence	Martin 1994
deps	any cycle	Dependency cycles between modules	Tarjan 1972
orphans	disabled	Unreferenced symbols (advisory)	—
class.coupling	CBO > 8	Classes coupled to too many other classes	Chidamber and Kemerer 1994
class.inheritance.depth	DIT > 4	Inheritance hierarchies that are too deep	Chidamber and Kemerer 1994
class.inheritance.children	NOC > 10	Base classes with too many direct subclasses	Chidamber and Kemerer 1994

Per-threshold explanations, when to raise them, and the default / lax / strict profiles live in the configuration reference.

Why a 14-day hotspot window

Tornhill's original work used a 1-year window, tuned for release cycles where a file you had not touched in 9 months was stable and a file you had been touching for 9 months was probably structurally important. Agent workflows collapse that timescale. A file can go from 200 LOC to 800 LOC in a week, and the architectural decisions compounding inside that growth are the ones worth catching early rather than a year later when the file is already unrecoverable. The 14-day default rewards recency. Widen rules.hotspots.since to "90 days ago" for human-pace repos.

Language support

Language	Extensions	Complexity	Hotspots	Packages	Deps	Class
Python	.py	yes	yes	yes	yes	yes
JavaScript	.js, .mjs, .cjs	yes	yes	—	yes	yes
TypeScript	.ts, .tsx	yes	yes	—	yes	yes
Go	.go	yes	yes	yes	yes	yes
Rust	.rs	yes	yes	—	—	yes
Java	.java	yes	yes	—	yes	yes
C#	.cs	yes	yes	—	yes	yes

Martin's package metrics currently only run on Go and Python. Files in other languages are silently excluded from that rule, not errored. Every other rule runs across all seven languages.

Configuration

slop walks upward from the current directory looking for .slop.toml first and then pyproject.toml with a [tool.slop] table, the same discovery ruff and mypy use. When a config is discovered, its root key resolves relative to the config file's directory, so root = "src" in ~/project/.slop.toml always points at ~/project/src regardless of which subdirectory you invoked slop from. --config and --root on the CLI override both.

Generate a starter config:

slop init          # balanced defaults
slop init lax      # legacy or gradual adoption
slop init strict   # greenfield or quality-focused

Every threshold, the intent behind it, and the three profiles in full are documented in the configuration reference.

CLI

slop lint                         Run all enabled rules (default command)
slop check <category|rule>        Run one category or rule
slop init [profile]               Generate .slop.toml
slop doctor                       Check fd, rg, git are installed
slop hook                         Install a git pre-commit hook
slop skill <dir>                  Install the bundled agent skill
slop rules                        List rules with thresholds
slop schema                       Config schema as JSON

Run slop --help for the full flag list, or slop <command> --help for per-command options. Output formats are --output human (default), --output json for CI and agent consumption, and --output quiet for one-line summaries.

Architecture

slop is a thin orchestration layer. Metric computation lives in aux-skills, a sibling package that ships deterministic kernels built on tree-sitter, ripgrep, fd, and git. slop wraps those kernels with declarative config, threshold checking, human/JSON/quiet output, and CI exit codes.

When aux-skills gains a new metric kernel, slop gains a new rule with zero computation work.

Acknowledgments

slop implements metrics from established software engineering research. Full citations are in NOTICE.

Metric	Author(s)	Year
Cyclomatic Complexity	Thomas J. McCabe	1976
Cognitive Complexity	G. Ann Campbell (SonarSource)	2018
CBO, DIT, NOC, WMC	Shyam R. Chidamber and Chris F. Kemerer	1994
Instability, Abstractness, D'	Robert C. Martin	1994, 2002
Hotspot analysis	Adam Tornhill	2015
Dependency cycle detection	Robert E. Tarjan	1972

These are mathematical formulas computed from source code structure. slop implements them independently via tree-sitter AST traversal. No code from the original authors' implementations is used.

License

Apache 2.0. See LICENSE.