Summary

Lingxi is an open-source, multi-agent framework designed for automated repository-level software issue resolution. It targets developers and researchers seeking to enhance automated program repair capabilities. By decomposing complex repair workflows into specialized agents and leveraging historical development knowledge, Lingxi significantly improves the accuracy and efficiency of bug fixing, achieving state-of-the-art performance on benchmarks like SWE-bench Verified.

How It Works

The framework employs a multi-agent architecture, with distinct agents for issue analysis (Problem Decoder), plan generation (Solution Mapper), code implementation (Problem Solver), and validation (Reviewer). Lingxi v1.5 mines transferable procedural knowledge from historical issue-patch pairs, injecting this prior into analysis agents. Lingxi v2.0 introduces a novel "trajectory-to-guidance" mechanism. This distills stage-aware procedural guidance from past repair traces, focusing on how to localize, validate, and iterate, rather than merely pattern matching. This approach aims to overcome context dilution and provide more adaptive, effective repair strategies.

Quick Start & Requirements

• Installation: Install via pip install -e . or uv sync within the cloned repository.
• Prerequisites: Requires a Python environment, LLM provider configuration (LLM_PROVIDER, LLM_MODEL), corresponding API keys (ANTHROPIC_API_KEY, OPENAI_API_KEY, DEEPSEEK_API_KEY), and a GITHUB_TOKEN. The search_relevant_files tool necessitates access to OpenAI's text-embedding-3-small model.
• Running: Execute langgraph dev --no-reload (pip) or uv run --env-file .env langgraph dev --no-reload (uv) to launch a local LangGraph Studio instance. Usage involves providing a GitHub issue URL.
• Resources: Links to official quick-start, docs, or demos are not explicitly provided in the README text.

Highlighted Details

• Lingxi v2.0 achieved 81.2% Pass@1 on SWE-bench Verified, ranking #1 on the leaderboard.
• Lingxi v1.5 achieved 74.6% Pass@1 on SWE-bench Verified using Claude 4 Sonnet.
• v2.0's trajectory-to-guidance mechanism provides stage-specific procedural insights.
• v1.5 leverages historical issue-patch data for knowledge-guided analysis.
• The system utilizes a "minimal tool set, maximal information" philosophy with structured function-calling.

Maintenance & Community

No specific details regarding maintainers, community channels (e.g., Discord, Slack), or public roadmaps are present in the provided README text.

Licensing & Compatibility

The README text does not specify the project's license or any compatibility notes for commercial use.

Limitations & Caveats

Runtime data (cloned repos, ChromaDB) is stored in a directory requiring manual cleanup for disk space management. The search_relevant_files tool has a direct dependency on OpenAI's embedding model. A publication for Lingxi v2.0 is noted as "coming soon," suggesting ongoing research and potential for further evolution.