
ContribFlow is a multi-agent pipeline built for the IBM Bob Hackathon that automates the hardest parts of open-source contribution. Stage 1 (Gap Finder) scans a GitHub repository using heuristic pre-scoring to identify meaningful contribution opportunities — missing error handling, unimplemented functions, deprecated dependencies — without wasting tokens on irrelevant files. Stage 2 (Idea Deduplication) uses local sentence-transformer embeddings and LLM verification to check whether your contribution idea already exists as an open issue or merged PR, preventing duplicate work. Stage 3 (Change Impact Analysis) is the core stage. It uses GitHub Code Search API combined with a LangGraph ReAct agent to map the full blast radius of a proposed change — identifying affected files, services at risk, and tests to update, with confidence-scored dependency traces. Validated against a real merged PR (opensre/tracer-cloud #1395) achieving 100% recall and 88.9% precision. Stage 4 (Pre-PR Quality Check) runs Ruff linting deterministically, samples repository conventions, and uses an LLM to catch issues requiring judgment before CI runs. All four stages are unified under a LangGraph orchestrator with real-time SSE streaming to a React frontend.
17 May 2026