RepoRevenant — Forgotten Code Deserves a Second Life

A developer built RepoRevenant, an open-source AI platform that diagnoses and repairs abandoned Python repositories by combining AST-based evidence, local open-weight LLMs (Llama 3, Mistral, CodeLlama via Ollama), and an ephemeral pytest sandbox that only declares a "resurrection" when post-patch tests exit with code 0. The pipeline enforces a strict epistemic hierarchy in which AI hypotheses and proposed diffs require human approval before verification, and it targets common breakages such as PEP 594 module removals and collections.abc drift. A live demo runs against two sample legacy repositories, with the full stack available locally via FastAPI and Next.js.

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine https://dev.to/challenges/hacktoberfest-week4-2026-10-26 RepoRevenant is an open-source, AI-powered software forensics and repository resurrection platform. "Forgotten code deserves a second life." Software projects are abandoned every day — not because their core ideas were wrong, but because language runtimes move on. A repository written in 2017 often fails silently on Python 3.12+ because of PEP 594 removals imp , cgi , pipes , collections.abc drift, and print-statement syntax errors. RepoRevenant enforces a strict epistemic hierarchy — the AI is never trusted blindly: | Tier | Source | Can Declare Resurrection? | |---|---|---| | Observed Facts — AST errors, dependency scans | Python ast , GitPython | No establishes baseline | | AI Hypotheses — root-cause diagnosis | Local Open-Weight LLM Llama 3, Mistral | No unverified | | Proposed Repairs — unified diffs | AI Agent Engine + Unified Patcher | No pending human approval | | Human Approval — developer consent | Interactive Review UI | Prerequisite | | Verified Outcome — post-patch test suite exit code 0 | Ephemeral Pytest Sandbox | YES — Empirical Proof | The theme "Ghost in the Machine" maps directly: the ghost is the dormant intelligence trapped inside an abandoned codebase. RepoRevenant is the medium that calls it back. 🔗 Live Demo : reporevenant.vercel.app https://reporevenant.vercel.app read-only sample repos The live deployment runs the frontend on Vercel proxied to a secured backend exposing two controlled sample repositories: sample-pyvault — a legacy secrets manager with PEP 594 imp and cgi violations sample-math-matrix — a 2016 linear algebra library with collections.Mapping drift Local full-power run: git clone https://github.com/divyanshigupta786/Repo-Revenant.git cd Repo-Revenant npm install python -m venv .venv && .venv\Scripts\activate pip install -r backend/requirements.txt npm run backend FastAPI on :8000 npm run dev Next.js on :3000 "Forgotten code deserves a second life." Investigate the forgotten. Repair the broken. Prove the resurrection. RepoRevenant is an open-source, AI-powered software investigation and repository resurrection platform built for the "Ghost in the Machine" open-source AI challenge. Software projects are frequently abandoned when maintainers move on, dependencies rot, or runtime platforms advance. A codebase written in 2017 often fails on modern runtimes—not because its core algorithms are flawed, but because language syntax shifted, standard library modules were purged such as Python's PEP 594 removal of imp , cgi , and pipes , and abstract base classes drifted. RepoRevenant inspects repository structure, extracts empirical Abstract Syntax Tree AST evidence, isolates the code in clean temporary sandboxes, executes baseline test suites to document exact pre-patch failures, synthesizes targeted repairs using replaceable open-weight AI models, and proves the resurrection by verifying post-patch test suites pass with exit code 0. Key source files: backend/app/main.py backend/app/engine/ backend/app/db.py components/HeroAnimations.jsx RepoRevenant is a 7-stage forensic pipeline built around fully local, replaceable open-weight models: 169.254.169.254 requirements.txt / setup.py , discovers test suites, scans language composition collections.abc drift, Python 2 print statements — pytest in an isolated temporary sandbox env-scrubbed, 30s timeout llama3:8b , mistral:7b , codellama:7b for contextual root-cause hypotheses and repair synthesis ast.parse , creates rollback backups, applies with human approval .md , telemetry as .json , diffs as .patch prefers-reduced-motion compliant | Model | Role | |---|---| | llama3:8b Ollama | Primary diagnosis + repair synthesis | | mistral:7b Ollama | Alternate repair synthesis | | codellama:7b Ollama | Code-specialized repair fallback | | Built-in Heuristic Expert | Air-gapped deterministic fallback zero neural API | If RepoRevenant were built on a closed API, it would be fundamentally unusable for its own purpose. Abandoned repositories often contain private intellectual property, proprietary algorithms, and trade secrets — the exact code an organization cannot afford to send to a third-party cloud endpoint for "repair." The ability to run llama3:8b locally via Ollama means: MODEL PROVIDER env var lets developers substitute any OpenAI-compatible local endpoint. Llama 3 today, a future specialized code-repair model tomorrow A closed API wouldn't just be a privacy risk — it would be a trust violation against the very repositories users are trying to resurrect. Session saved via DevRelay — link coming after final verification run. There's a particular kind of sadness to an abandoned repository. Someone cared enough to write it — to name it, to structure it, to leave comments explaining why. Then life happened: they moved on, the company pivoted, the project "finished." Five years later, python setup.py install explodes with ModuleNotFoundError: No module named 'imp' , and all that careful work becomes invisible. RepoRevenant started from a single observation: the code still knows what it wanted to do. The AST is still there. The test assertions still describe the intended behavior. The ghost is still in the machine. Watching the Three.js resurrection scene — fragmented archival nodes slowly pulling back together as the post-patch test suite returns exit code 0 — that's the moment it lands. Not because the AI did something magic, but because the code was never actually broken. It just needed someone to listen to it again.