Slopwatch

// LLM anti-cheat
A .NET tool that detects LLM "reward hacking" behaviors in code changes.

---

Runs as a Claude Code hook or in CI/CD pipelines to catch when AI coding assistants take shortcuts instead of properly fixing issues.

What is "Slop"?

When LLMs generate code, they sometimes take shortcuts that make tests pass or builds succeed without actually solving the underlying problem. These patterns include:

• Disabling tests instead of fixing them ([Fact(Skip="flaky")])
• Suppressing warnings instead of addressing them (#pragma warning disable)
• Swallowing exceptions with empty catch blocks
• Adding arbitrary delays to mask timing issues (Task.Delay(1000))
• Project-level warning suppression (<NoWarn>, <TreatWarningsAsErrors>false</TreatWarningsAsErrors>)
• Bypassing Central Package Management with VersionOverride or inline Version attributes
• And more...

Slopwatch catches these patterns before they make it into your codebase.

Installation

# Install as a global tool
dotnet tool install --global Slopwatch.Cmd

# Or install locally
dotnet tool install Slopwatch.Cmd

Quick Start

# 1. Install slopwatch
dotnet tool install --global Slopwatch.Cmd

# 2. Initialize in your project (creates baseline from existing code)
cd your-project
slopwatch init

# 3. Commit the baseline to your repository
git add .slopwatch/baseline.json
git commit -m "Add slopwatch baseline"

# 4. From now on, only NEW slop is detected
slopwatch analyze

The baseline approach ensures slopwatch catches new slop being introduced without flagging legacy code. Your CI/CD pipeline will fail if someone introduces new slop patterns.

Usage

Initialize a Project

# Create baseline from existing code
slopwatch init

# Force overwrite existing baseline
slopwatch init --force

This creates .slopwatch/baseline.json containing all existing detections. Commit this file to your repository.

Analyze for New Issues

# Analyze current directory (requires baseline by default)
slopwatch analyze

# Analyze specific directory
slopwatch analyze -d src/

# Skip baseline and report ALL issues (not recommended for CI)
slopwatch analyze --no-baseline

Update Baseline

When you intentionally add code that triggers slopwatch (with proper justification):

# Add new detections to existing baseline
slopwatch analyze --update-baseline

Output Formats

# Human-readable console output (default)
slopwatch analyze

# JSON for programmatic use
slopwatch analyze --output json

Exit Codes

# Fail if errors found (default)
slopwatch analyze --fail-on error

# Fail on warnings too
slopwatch analyze --fail-on warning

Performance

If you're concerned about performance on large projects, use --stats to see how many files are being analyzed and how long it takes:

slopwatch analyze --stats
# Output: Stats: 44 files analyzed in 1.64s

Detection Rules

Rule	Severity	Description
SW001	Error	Disabled tests via Skip, Ignore, or #if false
SW002	Warning	Warning suppression via pragma or SuppressMessage
SW003	Error	Empty catch blocks that swallow exceptions
SW004	Warning	Arbitrary delays in test code (Task.Delay, Thread.Sleep)
SW005	Warning	Project file slop (NoWarn, TreatWarningsAsErrors=false)
SW006	Warning	CPM bypass via VersionOverride or inline Version attributes

Claude Code Integration

Add slopwatch as a hook to catch slop patterns during AI-assisted coding. Add the following to your project's .claude/settings.json:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Write|Edit|MultiEdit",
        "hooks": [
          {
            "type": "command",
            "command": "slopwatch analyze -d . --hook",
            "timeout": 60000
          }
        ]
      }
    ]
  }
}

The --hook flag enables Claude Code integration mode which:

• Uses git status to only analyze dirty files - makes hooks near-instant even on large repos
• Outputs errors to stderr in a readable format
• Suppresses all other output
• Blocks on warnings and errors
• Exits with code 2 on failure (blocking the edit)

Claude will see the formatted error message and can then fix the issue properly.

Note: Hook mode requires git. If git is unavailable, it falls back to full analysis.

CI/CD Integration

GitHub Actions

- name: Install Slopwatch
  run: dotnet tool install --global Slopwatch.Cmd

- name: Run Slopwatch
  run: slopwatch analyze -d . --output json --fail-on error

Note: The baseline file (.slopwatch/baseline.json) should be committed to your repository. Run slopwatch init locally first.

Azure DevOps

- task: DotNetCoreCLI@2
  displayName: 'Install Slopwatch'
  inputs:
    command: 'custom'
    custom: 'tool'
    arguments: 'install --global Slopwatch.Cmd'

- script: slopwatch analyze -d . --fail-on error
  displayName: 'Run Slopwatch'

Configuration

Create a .slopwatch/slopwatch.json configuration file to customize behavior:

{
  "minSeverity": "warning",
  "rules": {
    "SW001": { "enabled": true, "severity": "error" },
    "SW002": { "enabled": true, "severity": "warning" },
    "SW003": { "enabled": true, "severity": "error" },
    "SW004": { "enabled": true, "severity": "warning" },
    "SW005": { "enabled": true, "severity": "warning" },
    "SW006": { "enabled": true, "severity": "warning" }
  },
  "exclude": ["**/Generated/**", "**/obj/**", "**/bin/**"]
}

Use the -c or --config option to specify a custom configuration file location:

slopwatch analyze -d . --config path/to/config.json

Building from Source

dotnet build
dotnet test
dotnet pack

Contributing

1. Fork and create a feature branch
2. Make changes and add tests
3. Submit a pull request

Note: This project uses slopwatch on itself - your PR will be analyzed for slop patterns!

License

Apache 2.0 - see LICENSE for details.

Inspiration

• Slopometry - Python equivalent for Claude Code
• Incrementalist - Git diff analysis patterns