Cybermule is an autonomous AI-powered CLI agent that detects test failures, analyzes them using LLMs (Claude, OpenAI), and automatically proposes, applies, verifies, and tracks code fixes — safely and iteratively, using Git.

🔧 Designed to fix failing tests, review commits, and track retry attempts in a memory graph. Modular, pluggable, and optimized for LLM-assisted debugging.

aider only work with python 3.12 in mac

Install required dependencies:

# Install system dependencies
brew install --HEAD universal-ctags/universal-ctags/universal-ctags  # macOS
sudo apt install universal-ctags  # Ubuntu

# Install Aider
pip install aider-chat

# Clone and install Cybermule
git clone https://github.com/etherzhhb/cybermule.git
cd cybermule
pip install -e .

Here's a sample configuration for Cybermule using LiteLLM:

litellm:
  model: "claude-3-5-haiku-20241022"
  max_tokens: 8192
  debug_prompt: true

setup_command: |
  pip install -e .

test_command: |
  pytest --tb=short -q

single_test_command: |
  pytest -q -k {test_name} --tb=short

Cybermule automates the test-debug-fix cycle using this loop:

1. Run Pytest → detect failing tests
2. Extract First Failure → parse traceback
3. LLM Summary → explain failure in natural language
4. Fix Plan Generation → propose JSON fix plan
5. Code Edit (via Aider) → patch the code with Git safety
6. Rerun Test → validate fix or retry with new branch
7. Track in Memory Graph → store attempt + reasoning trace

cybermule/
├── commands/            # CLI entrypoints (e.g., run-and-fix)
├── executors/           # LLM logic: traceback analysis, fix generation
├── tools/               # Git and test runners
├── utils/               # Symbol resolution, AST tools
├── memory/              # Memory graph and commit history
├── prompts/             # Prompt templates (Jinja2)
├── symbol_resolution/   # Ctags + AST + semantic search

# Run the full fix loop
cybermule run-and-fix

# Review last commit
cybermule review-commit

Supports filtering by test name, file, or error.

{
  "fix_description": "Fix condition in test case",
  "symbol": "TestExample.test_something",
  "edits": [
    {
      "file": "tests/test_example.py",
      "line": 42,
      "code_snippet": "assert x == 1"
    }
  ]
}

All reasoning steps (traceback → summary → fix plan) are stored in a DAG using MemoryGraph, enabling future review, debugging, or replay.

Also used for:

• Retry tracking
• Fix lineage
• Commit review prompts

When the LLM says it needs more code context:

{
  "needs_more_context": true,
  "request_description": "...",
  "required_info": [
    {
      "symbol": "extract_locations",
      "ref_path": "utils/context_extract.py",
      "ref_function": "get_context_snippets"
    }
  ]
}

Cybermule extracts relevant snippets using:

• AST (preferred)
• Ctags
• Tree-sitter fallback
• Optional semantic search (e.g. FAISS)

• Claude via AWS Bedrock
• OpenAI (GPT-4)
• Ollama (local inference)

pytest tests/

Includes tests for:

• CLI
• Fix generation
• Symbol resolution
• Git safety
• Context enrichment

• design.md – System architecture
• implementation_plan.md – Development roadmap

• 🔒 Safe by Default – Git branches, commits, undo support
• 🔁 Fix, Test, Retry – Multi-round AI reasoning
• 🧠 Context-Aware – Extracts only what the LLM needs
• 💬 Transparent – All decisions logged and reviewable

Pull requests and issues welcome!