VitAI & GlitchWitcher: Repository analysis and defect prediction

01 / Problem & ownership

Context retrieval for repository investigation and ML-based defect assessment.

02 / How it works

MCP repository intelligence and automated PR defect analysis.

VitAI: An MCP interface exposes repository-analysis capabilities to compatible clients. A ReAct agent can explore repository structure, search GitHub, retrieve file contents, and query indexed code before answering a question. Live GitHub exploration and vector retrieval serve different purposes: live tools inspect repository content and discussions; the retrieval layer searches pre-indexed chunks for relevant context.

GlitchWitcher: The workflow extracts software metrics from changed C/C++ files and uses a trained REPD model to compare base and head predictions. Analysis can be triggered through a pull-request or issue comment, with results returned to the discussion.

03 / Engineering decisions

Structural chunking and research implementation with workflow integration.

Syntax-aware retrieval: VitAI’s indexing layer uses Tree-sitter-based chunking with fallback strategies, FAISS storage, and incremental change detection. This supports retrieval around code structure rather than treating a repository as an undifferentiated text document.

Research implementation with workflow integration: GlitchWitcher implements REPD from published research. Reconstruction errors are compared against fitted distributions for historically defective and non-defective examples.

File-level comparisons: The reported output helps inspect changes between a pull request’s base and head. It is not a claim that individual faulty lines have been located or that a file has a calibrated percentage probability of containing a bug.

# Simplified pseudocode
from fastmcp import FastMCP
from agent import create_agent

mcp = FastMCP("VitAI MCP Server")

@mcp.tool
def query(input: str) -> str:
    """Ask VitAI questions related to Adoptium repositories."""
    agent = get_agent()
    return agent.query(input)

@mcp.tool
def query_vector_store(query: str, repository: str, top_k: int = 5) -> str:
    """Query pre-indexed vector stores for semantic code search across repositories."""
    store = get_vector_store()
    return store.search_repo(query, repository, top_k=top_k)

04 / Evidence & limitations

Upstream contributions and model assessment boundaries.

The work produced upstream contributions, including VitAI’s initial retrieval components and GlitchWitcher’s GitHub Actions workflow.

Prediction outputs are model assessments, not proof that a file is defective or defect-free. GlitchWitcher reports probability-density values, not calibrated bug probabilities. I do not present the implementation as an independently measured reduction in production defects.