slopforge

Detect and clean up AI slop in your codebase — dead code, duplication, over-engineered abstractions, inconsistent style, unused dependencies, dead files.

slopforge is two coordinated surfaces sharing one core:

1. A CLI (npx slopforge) that runs the deterministic pipeline: it shells out to mature existing tools (Knip, Biome/Prettier, jscpd, Ruff, Vulture, deptry), normalizes their output into one unified report, and applies fixes in safety tiers behind git checkpoints.
2. A Claude Code skill (slopforge audit) that runs the judgment pipeline: over-engineering detection and semantic (non-token-identical) duplication — the lanes no static tool can own — as a bounded, verify-driven agent loop.

The CLI is the hands; the skill is the judgment. The CLI is fully usable standalone (CI, no agent).

Quick start

# 1. Analyze — writes .slopforge/report.md, modifies nothing
npx slopforge scan

# 2. Review — open the report in any editor, check the boxes you approve
code .slopforge/report.md   # or: open / vim / your editor of choice

# 3. Apply — git checkpoint, then tiered execution with test gating
npx slopforge apply

# Changed your mind?
npx slopforge revert --last

Design principles

• Wrap, don't reimplement. Every deterministic lane delegates to the best existing tool. slopforge's value is orchestration, one unified report, safe tiered auto-fix, and the agent pass — not novel static analysis.
• Report before write, always. Nothing is modified until you've approved it in the report. Even "safe" fixes are gated behind report approval.
• Tiered execution behind git. Every apply starts with a git checkpoint. Tier 1 (formatting, safe lint fixes) is pre-approved in the report; tier 2 (deletions, dependency removal) is opt-in per item with test runs between batches; tier 3 (judgment refactors) is agent-only and always diff-reviewed.
• Language-pluggable. v1 ships JavaScript/TypeScript and Python. Adding a language is one LanguageAdapter, not a core change.
• False-positive-aware. Dead-code tools lie (dynamic imports, reflection, DI, framework magic). Nothing destructive is ever auto-applied.

Commands

| Command | What it does | |---|---| | slopforge scan [path] | Run all analyzers, write .slopforge/report.md + findings.json. Read-only. | | slopforge apply [path] | Execute the findings you approved in the report, behind a checkpoint. | | slopforge revert [path] | Roll back to a checkpoint (--last or --all). | | slopforge prevent [path] | Install a short anti-slop ruleset into the repo's agent config (asks first). | | slopforge audit … | Harness subcommands for the agent judgment loop (driven by the skill). |

Documentation

• Getting started — install, first scan, first apply
• Concepts & architecture — lanes, tiers, findings, adapters, the report
• CLI reference — every command and flag
• Configuration — slopforge.config.json
• The audit loop — the agent judgment pass: verify gates, ledger, maker/checker
• Worked example — a full end-to-end run on a real GitHub repo

Requirements

• Node.js ≥ 20 and git.
• The underlying analyzers for the languages you use. slopforge detects what's installed, degrades gracefully, and prints exact install hints for anything missing — it never installs tools itself.

License

MIT