dotnet-slopwatch
Use Slopwatch to detect LLM reward hacking in .NET code changes. Run after every code modification to catch disabled tests, suppressed warnings, empty catch blocks, and other shortcuts that mask real problems.
- License 许可
- MIT
- 最近更新
- 2026-11-13
- Stars
- ★ 1
- Forks
- 0
01 前言 Front Matter
- name
- dotnet-slopwatch
- description
- Use Slopwatch to detect LLM reward hacking in .NET code changes. Run after every code modification to catch disabled tests, suppressed warnings, empty catch blocks, and other shortcuts that mask real problems.
03 SKILL.md 正文
Slopwatch: LLM Anti-Cheat for .NET
When to Use This Skill
Use this skill constantly. Every time an LLM (including Claude) makes changes to:
- C# source files (.cs)
- Project files (.csproj)
- Props files (Directory.Build.props, Directory.Packages.props)
- Test files
Run slopwatch to validate the changes don't introduce "slop."
What is Slop?
"Slop" refers to shortcuts LLMs take that make tests pass or builds succeed without actually solving the underlying problem. These are reward hacking behaviors - the LLM optimizes for apparent success rather than real fixes.
Common Slop Patterns
| Pattern | Example | Why It's Bad |
|---|---|---|
| Disabled tests | [Fact(Skip="flaky")] |
Hides failures instead of fixing them |
| Warning suppression | #pragma warning disable CS8618 |
Silences compiler without fixing issue |
| Empty catch blocks | catch (Exception) { } |
Swallows errors, hides bugs |
| Arbitrary delays | await Task.Delay(1000); |
Masks race conditions, makes tests slow |
| Project-level suppression | <NoWarn>CS1591</NoWarn> |
Disables warnings project-wide |
| CPM bypass | Version="1.0.0" inline |
Undermines central package management |
Never accept these patterns. If an LLM introduces slop, reject the change and require a proper fix.
Installation
Ask the user before you install slopwatch, create or update its baseline, or commit the baseline or
the tool manifest: each changes the repository or the machine. When the tool or the baseline is
missing, say so and wait for the answer; running slopwatch analyze against an existing setup needs
no question.
As a Local Tool (Recommended)
Add to .config/dotnet-tools.json:
{
"version": 1,
"isRoot": true,
"tools": {
"slopwatch.cmd": {
"version": "0.4.2",
"commands": ["slopwatch"],
"rollForward": false
}
}
}
Then restore:
dotnet tool restore
As a Global Tool
dotnet tool install --global Slopwatch.Cmd
First-Time Setup: Establish a Baseline
Before using slopwatch on an existing project, create a baseline of current issues:
# Initialize baseline from existing code
slopwatch init
# This creates .slopwatch/baseline.json
git add .slopwatch/baseline.json
git commit -m "Add slopwatch baseline"
Why baseline? Legacy code may have existing issues. The baseline ensures slopwatch only catches new slop being introduced, not pre-existing technical debt.
Usage During LLM Sessions
After Every Code Change
Run slopwatch after any LLM-generated code modification:
# Analyze for new issues (uses baseline)
slopwatch analyze
# Use strict mode - fail on warnings too
slopwatch analyze --fail-on warning
When Slopwatch Flags an Issue
Do not ignore it. Instead:
- Understand why the LLM took the shortcut
- Request a proper fix - be specific about what's wrong
- Verify the fix doesn't introduce different slop
# Example: LLM disabled a test
❌ SW001 [Error]: Disabled test detected
File: tests/MyApp.Tests/OrderTests.cs:45
Pattern: [Fact(Skip="Test is flaky")]
# Correct response: Ask for actual fix
"This test was disabled instead of fixed. Please investigate why
it's flaky and fix the underlying timing/race condition issue."
Updating the Baseline (Rare)
Only update the baseline when slop is truly justified and documented:
# Add current detections to baseline (use sparingly!)
slopwatch analyze --update-baseline
Justification examples:
- Third-party library forces a pattern (e.g., must suppress specific warning)
- Intentional delay for rate limiting (not test flakiness)
- Generated code that can't be modified
Document why in a code comment when updating baseline.
Claude Code Hook Integration
Add slopwatch as a hook to automatically validate every edit. Create or update .claude/settings.json:
{
"hooks": {
"PostToolUse": [
{
"matcher": "Write|Edit|MultiEdit",
"hooks": [
{
"type": "command",
"command": "slopwatch analyze -d . --hook",
"timeout": 60000
}
]
}
]
}
}
The --hook flag:
- Only analyzes git dirty files (fast, even on large repos)
- Outputs errors to stderr in readable format
- Blocks the edit on warnings/errors (exit code 2)
- Claude sees the error and can fix it immediately
CI/CD Integration
Add slopwatch to your CI pipeline as a quality gate:
GitHub Actions
jobs:
slopwatch:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- name: Setup .NET
uses: actions/setup-dotnet@v4
with:
dotnet-version: '9.0.x'
- name: Install Slopwatch
run: dotnet tool install --global Slopwatch.Cmd
- name: Run Slopwatch
run: slopwatch analyze -d . --fail-on warning
Azure Pipelines
- task: DotNetCoreCLI@2
displayName: 'Install Slopwatch'
inputs:
command: 'custom'
custom: 'tool'
arguments: 'install --global Slopwatch.Cmd'
- script: slopwatch analyze -d . --fail-on warning
displayName: 'Slopwatch Analysis'
Detection Rules
| Rule | Severity | What It Catches |
|---|---|---|
| SW001 | Error | Disabled tests (Skip=, Ignore, #if false) |
| SW002 | Warning | Warning suppression (#pragma warning disable, SuppressMessage) |
| SW003 | Error | Empty catch blocks that swallow exceptions |
| SW004 | Warning | Arbitrary delays in tests (Task.Delay, Thread.Sleep) |
| SW005 | Warning | Project file slop (NoWarn, TreatWarningsAsErrors=false) |
| SW006 | Error | CPM bypass (VersionOverride, inline Version attributes) |
Configuration
Create .slopwatch/config.json to suppress a rule for specific paths, with a justification:
{
"suppressions": [
{
"ruleId": "SW002",
"pattern": "**/Generated/**",
"justification": "Generated code from tooling cannot be manually changed"
}
],
"globalSuppressions": []
}
Use -c or --config to point at a config file elsewhere, and --exclude to leave paths out of analysis entirely:
slopwatch analyze -d . --config path/to/config.json
slopwatch analyze -d . --exclude "**/Generated/**,**/obj/**,**/bin/**"
Strict Mode (Recommended for LLM Sessions)
For maximum protection during LLM coding sessions, fail on warnings as well as errors:
slopwatch analyze -d . --fail-on warning
The Philosophy: Zero Tolerance for New Slop
- Baseline captures legacy - Existing issues are acknowledged but isolated
- New slop is blocked - Any new shortcut fails the build/edit
- Exceptions require justification - If you must update baseline, document why
- LLMs are not special - The same rules apply to human and AI-generated code
The goal is to prevent the gradual accumulation of technical debt that occurs when LLMs optimize for "make the test pass" rather than "fix the actual problem."
Quick Reference
# First time setup
slopwatch init
git add .slopwatch/baseline.json
# After every LLM code change
slopwatch analyze
# Strict mode (recommended)
slopwatch analyze --fail-on warning
# With stats (performance debugging)
slopwatch analyze --stats
# Update baseline (rare, document why)
slopwatch analyze --update-baseline
# JSON output for tooling
slopwatch analyze --output json
When to Override (Almost Never)
The only valid reasons to update baseline or disable a rule:
| Scenario | Action | Required |
|---|---|---|
| Third-party forces pattern | Update baseline | Code comment explaining why |
| Generated code (not editable) | Add to exclude list | Document in config |
| Intentional rate limiting delay | Update baseline | Code comment, not in test |
| Legacy code cleanup | One-time baseline update | PR description |
Invalid reasons:
- "The test is flaky" → Fix the flakiness
- "The warning is annoying" → Fix the code
- "It works on my machine" → Fix the race condition
- "We'll fix it later" → Fix it now
04 同仓库更多 Skills
writing-tunit-tests
Write tests in TUnit — the attributes, the async assertion model, data-driven cases, lifecycle hooks, and the parallelism rules. USE FOR writing or fixing a test in a TUnit project, translating a test idea into TUnit syntax, or reviewing TUnit test code for API mistakes. DO NOT USE FOR deciding what to test, judging whether a test is any good, or running tests.
dotnet-walls
Walls for a .NET change in a repository without its own quality gates — pinned, ephemeral analyzer, format, clone and scanner commands for the verify step, their identity ratchet, and the build traps. Use during a walls preflight, or when a worker's verify step needs .NET quality gates.
csharp-15
Write, review and modernize C# 15 (.NET 11) code correctly. Covers every C# 15 feature - collection expression arguments (`[with(...), ...]`), union types (`union`, `[Union]`, `IUnion`, custom unions), closed class hierarchies (`closed`), extension indexers, labeled `break`/`continue`, non-virtual static interface members on runtimes without default interface methods, and the preview memory-safety model (`unsafe(expr)`, `safe`, requires-unsafe members). Use when code targets net11.0 or LangVersion 15/preview, when the user mentions any of these features, when replacing marker interfaces, OneOf-style results, `goto`/flag loop exits or capacity/comparer boilerplate, or when the build reports CS8121, CS8509, CS8655, CS8510, CS0121/CS0411 on a `with(...)` element, CS9370, CS9380 or CS9363.
verified-waves
Run one task as verified multi-agent waves. Plans the fan-out, sends findings to fresh skeptics, keeps a ledger, and reports N of M done; small tasks stay solo.
harness
The bundled dispatch harness, which runs a task file on worker models through OpenCode or Codex, fences their writes, escalates their asks to the operator, settles each result from git and its verify commands, and records every step in a ledger. Its commands, task file, modes and write fence, configuration, runs and their recovery, ledger and exit codes.
delegating-to-worker-models
Use when you hand off implementation, a plan, research a decision will rely on, or a review, and you brief the worker and check what comes back. A subagent that searches, reads in parallel, or keeps your context small is ordinary tool use. Covers what stays with you, the mechanism, assigning work by strength, the walls preflight, prompts without contradictions, the worker contract and model profiles, evidence from your own run, review, and the fix-up loop.