Frequently Asked Questions (FAQ)

January 6, 2026 · View on GitHub

Getting Started

What is SlopCodeBench?

SlopCodeBench measures code erosion under iterative specification refinement — quantifying how code quality degrades as AI agents iterate on programming problems. It helps evaluate how agents handle evolving requirements.

What are the system requirements?

  • Python 3.12+
  • Docker (installed and running)
  • 8GB+ RAM recommended
  • 10GB+ disk space for Docker images and workspaces
  • API keys for the agents you want to test

How long does the first run take?

First run takes 5-10 minutes because Docker images need to build. Subsequent runs are much faster (typically 1-2 minutes per problem depending on the agent and model).

Do I need Docker?

Yes, Docker is required for isolated execution environments. This ensures reproducibility and prevents agents from affecting your local system.

Running Benchmarks

What does thinking=low|medium|high mean?

This controls the extended thinking token budget for models that support extended thinking:

  • low = 10,000 tokens
  • medium = 20,000 tokens
  • high = 40,000 tokens

What does the version parameter do?

Specifies the agent version to use (e.g., version=2.0.51). If omitted, the latest version is used. This is useful for reproducing results or testing specific agent versions.

Can I run multiple problems at once?

Yes, you can specify multiple --problem flags:

uv run slop-code run --problem file_backup --problem execution_server

How do I run with a different agent?

Use the --agent flag with one of the supported agents:

  • claude_code
  • codex
  • gemini
  • miniswe
  • opencode
  • openhands

See Agent Guide for configuration details.

Where are results saved?

Results are saved to outputs/{model_name}/{agent}-{prompt}_{params}_{timestamp}/

Each run creates:

  • results.json - Evaluation results
  • overall_quality.json - Code quality metrics
  • submissions/ - Agent-generated code
  • workspaces/ - Execution environments

Evaluation

What's the difference between correctness and quality metrics?

  • Correctness: Does the solution pass the test cases? (Pass/Fail)
  • Quality: How clean, maintainable, and well-structured is the code? (Metrics like complexity, duplication, etc.)

How do I interpret pass policies?

Pass policies determine what counts as "passing" an evaluation:

  • ALL_CASES - All test cases must pass
  • ANY_CASE - At least one test case must pass
  • MAJORITY - More than 50% of test cases must pass

What are checkpoints?

Checkpoints represent stages of specification refinement. Problems have multiple checkpoints, each adding requirements or complexity. This tests how code quality evolves as specifications change.

Can I create custom test cases?

Yes! See the Problem Tutorial for a step-by-step guide on creating problems with custom test cases.

Problems

How do I create a new problem?

Follow the Problem Tutorial which walks through creating a complete problem in ~30 minutes. Also see the Contributing Guide.

What makes a good problem?

Good problems:

  • Take ~40 hours to solve well (not just make it work)
  • Have clear, incremental checkpoints
  • Test realistic refactoring scenarios
  • Encourage flexible, maintainable code over quick hacks

See Problem Design Guide for details.

Agents

Which agent should I use?

It depends on your use case:

  • Claude Code: Best for complex refactoring, supports extended thinking
  • Codex: OpenAI's code model
  • Gemini: Google's model with long context
  • MiniSWE: Lightweight agent for quick testing
  • OpenCode: DeepSeek-based agent
  • OpenHands: Multi-tool agent

See agent-specific docs in docs/agents/agents/.

How do I configure API keys?

Set environment variables for your provider:

export ANTHROPIC_API_KEY="your-key"
export OPENAI_API_KEY="your-key"
export GOOGLE_API_KEY="your-key"

Or see Credentials Guide for file-based configuration.

Can I add a new agent?

Yes! See the Agent Implementation Guide and Contributing Guide.

Troubleshooting

Docker build fails

# Clean Docker cache and rebuild
docker system prune -a
uv run slop-code docker build-base --environment configs/environments/docker-python3.12-uv.yaml

"Command not found: docker"

Docker is not running or not installed. Install Docker Desktop and ensure it's running:

docker ps  # Should list running containers

API key errors

Verify your environment variable is set:

echo $ANTHROPIC_API_KEY  # Should print your key

Out of disk space

Docker images can take significant space. Clean up:

docker system prune -a  # Remove unused images
docker volume prune     # Remove unused volumes

Tests fail but solution looks correct

Check the expected outputs in the problem's test cases. Some test failures may be due to:

  • Formatting differences (whitespace, newlines)
  • Numeric precision (for floating point)
  • Output order (for unordered results)

See Troubleshooting Guide for handling edge cases.

Metrics

What metrics are collected?

  • Correctness: Pass/fail for each test case
  • Complexity: Cyclomatic complexity, nesting depth
  • Code Quality: Duplication, code smells, maintainability
  • Size: Lines of code, file count
  • Deltas: Changes between checkpoints

See Metrics Guide for complete list.

How do I run the LLM judge?

slop-code metrics judge \
  --rubric configs/rubrics/llm_judge.jsonl \
  --model <model on openrouter> \
  --criteria-template configs/rubrics/templates/criteria_with_pn.j2

What's a good score?

It depends on the problem and checkpoint. See Interpreting Results for guidance on what different metric values mean.

Contributing

How do I contribute a problem?

  1. Design the problem (see Design Philosophy)
  2. Write pytest tests (see Tutorial)
  3. Test thoroughly (see Checklist)
  4. Submit a PR (see Contributing Guide)

How long does PR review take?

This is early-stage software. Review times vary. We'll work with you to iterate on the problem design.

Can I contribute to documentation?

Absolutely! Documentation improvements are very welcome. See the Contributing Guide.

Advanced

Can I run evaluations in parallel?

Yes, you can run evaluations with multiple workers:

slop-code eval outputs/my_run --num-workers 4

This speeds up evaluation by running multiple problems or checkpoints in parallel.

How do I create custom metrics?

Custom metrics require modifying the metrics system. This is not yet documented. See existing metrics in src/slop_code/metrics/ for examples.

Can I use this in CI/CD?

Yes, you can integrate SlopCodeBench into your CI/CD pipeline. Results are saved in JSON format for easy parsing.

How do I reproduce results from a paper?

Use the same:

  • Agent version (via version=X.Y.Z)
  • Model
  • Environment configuration
  • Problem set
  • Random seed (if applicable)

Still have questions?