SWE-Debate, a competitive multi-agent debate framework that promotes diverse reasoning paths and achieves more consolidated fault localization.SWE-Debate reframes issue resolution through graph-guided local-ization and structured debate mechanisms.
This project integrates an advanced Entity Localization Pipeline based on the Moatless framework, achieving the following breakthrough features:
- 🔍 Intelligent Entity Extraction: Automatically identifies key code entities (classes, methods, functions, variables) from issue descriptions
- 🕸️ Graph-Driven Search: Performs deep traversal and relevance analysis based on code dependency graphs
- ⛓️ Localization Chain Generation: Constructs complete code localization chains from problems to solutions
- 🤖 Multi-Agent Collaborative Debate: Multiple expert agents collaborate in analysis and debate to ensure accuracy and comprehensiveness of solutions
- 📋 Automated Solution Generation: Generates detailed, highly instructive code modification guidance plans
Figure: Complete workflow of SWE-Debate
- Python 3.12+
- OpenAI API key or LLM API supporting OpenAI format
-
Clone the project
-
Install dependencies
# Install main dependencies
pip install -r localization/requirements.txt
# Install SWE-Search dependencies
pip install moatless-tree-search- Configure environment variables
# Create .env file
cp .env.example .env
# Edit .env file, add your API configuration
OPENAI_API_KEY=your_api_key_here
OPENAI_BASE_URL=your_base_url_here # Optional, for custom API endpointsfrom localization.entity_localization_pipeline import EntityLocalizationPipeline
# initialize
pipeline = EntityLocalizationPipeline(
model_name="deepseek/deepseek-chat",
max_depth=5
)
# run
result = pipeline.run_pipeline(
instance_data=instance_data,
context=file_context,
max_initial_entities=5
)react_completion_model = ReActCompletionModel(model="deepseek/deepseek-chat", temperature=0.7)
completion_model = CompletionModel(model="deepseek/deepseek-chat", temperature=0.7)
discriminator_model = CompletionModel(model="deepseek/deepseek-chat", temperature=1)
value_model = CompletionModel(model="deepseek/deepseek-chat", temperature=0.2)
react_completion_model.response_format = LLMResponseFormat.REACT
completion_model.response_format = LLMResponseFormat.REACT
discriminator_model.response_format = LLMResponseFormat.REACT
value_model.response_format = LLMResponseFormat.REACT
instance = get_moatless_instance(instance_id=instance_id)
repository = create_repository(instance)
index_store_dir = os.getenv("INDEX_STORE_DIR", os.path.abspath("tmp/index_store"))
code_index = CodeIndex.from_index_name(
instance["instance_id"], file_repo=repository, index_store_dir=index_store_dir
)
file_context = FileContext(repo=repository)
current_date = datetime.now().strftime("%Y-%m-%d")
instance_path = f'{base_dir}/trajectory/{instance_id}/'
persist_path = f'{base_dir}/trajectory/{instance_id}/{current_date}_trajectory.json'
value_function = ValueFunction(completion_model=value_model)
actions = [
FindClass(completion_model=react_completion_model, code_index=code_index, repository=repository),
FindFunction(completion_model=react_completion_model, code_index=code_index, repository=repository),
FindCodeSnippet(completion_model=react_completion_model, code_index=code_index, repository=repository),
SemanticSearch(completion_model=react_completion_model, code_index=code_index, repository=repository),
ViewCode(completion_model=react_completion_model, repository=repository),
StringReplace(repository=repository, code_index=code_index),
CreateFile(repository=repository, code_index=code_index),
Finish(),
]
system_prompt = AGENT_ROLE
if completion_model.response_format == LLMResponseFormat.REACT:
system_prompt += REACT_CORE_OPERATION_RULES
elif completion_model.response_format == LLMResponseFormat.TOOLS:
system_prompt += REACT_GUIDELINES
workflow_prompt = generate_workflow_prompt(actions, has_runtime=True)
system_prompt += workflow_prompt + generate_guideline_prompt(has_runtime=True) + ADDITIONAL_NOTES
agent = CodingAgent(system_prompt=system_prompt, actions=actions, completion=react_completion_model)
discriminator = AgentDiscriminator(
completion=discriminator_model,
n_agents=5,
n_rounds=3,
)
feedback_generator = FeedbackAgent(
completion_model=completion_model, instance_dir=instance_path
)
from entity_localization_pipeline import EntityLocalizationPipeline
pipeline = EntityLocalizationPipeline()
results = pipeline.run_pipeline(
instance,"context",
max_initial_entities=5
)
search_tree = SearchTree.create(
message=f'{results}',
agent=agent,
file_context=file_context,
value_function=value_function,
discriminator=discriminator,
feedback_generator=feedback_generator,
max_finished_nodes=max_finish_nodes,
max_iterations=max_iterations,
max_expansions=3,
max_depth=20,
max_duplicate_count=5,
persist_path=persist_path,
)
finished_node = search_tree.run_search()
search_tree.persist(persist_path)This project draws inspiration from and builds upon the methodologies presented in LocAgent. We acknowledge their contributions to the field of automated code localization.