This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine
What I Built
What I Built
Some ghosts rattle chains. Mine points at a commit.
A repository can tell you what its code looks like today. Understanding how it got there is harder.
Why was that workaround added? What happened to a deleted helper? Was that strange fallback meant to be permanent—or did “temporary” quietly become part of the architecture?
I built Git Séance, a local-first application that helps developers investigate those questions using Git history and a local Gemma model through Ollama.
The idea is simple:
Ask the ghost of your repository about its past, then inspect the commits behind its answer.
Git Séance brings together:
Repository questions: ask about filenames, symbols, and historical changes.
Commit-grounded explanations: relevant history becomes evidence for the model.
Inspectable sources: open commit metadata and patch excerpts instead of trusting an answer blindly.
A code graveyard: explore files that disappeared from the repository.
Creative epitaphs: give deleted code a farewell, clearly separated from factual history.
A searchable memory archive: browse commit subjects, authors, and hashes.
Demo
How it fits the theme
My interpretation of Ghost in the Machine is not that the AI knows everything.
It is that the machine still carries traces of what came before: removed files, abandoned approaches, temporary fixes, and decisions recorded in commits.
The graveyard is a small tribute to code that once served a purpose.
A deleted file is not necessarily bad code. Sometimes it completed its job.
Code
Code
GitHub: [https://github.com/Ganeshpatel404/git-seance.git]
The application code is released under the MIT License. Model weights remain subject to their own licence terms.
Run locally on Windows
Requirements:
Node.js 20 or newer
Git
Ollama
Install the generation model:
PowerShell
From the project folder:
PowerShell
Open:
text
Then connect a local repository that already has Git commits.
Semantic retrieval is optional. The application also supports keyword retrieval, so exploring history does not depend on downloading an embedding model.
Local repository access should not be enabled on a public deployment.
My Agent Session
Building with AI assistance
I used an AI coding assistant during development to help scaffold and iterate on the application, tests, and documentation.
[Add an accurate sentence about your own work: which design decisions you made, which outputs you reviewed, and what you tested yourself.]
I wanted to be explicit about that process rather than present generated code as unaided work.
Current limitations
The current version is intentionally scoped:
History scanning is bounded, not unlimited.
Tombstones represent deleted files, not comprehensive function-level deletion analysis.
Citation checking does not automatically establish factual support.
Unrecorded motives remain unknown.
History and semantic vectors are held in memory in the current implementation.
Large repositories and complex refactors need further evaluation.
[Update this list if your final implementation changed these behaviours.]
The Spirit of It
Prize Categories
Prize Categories
Best Use of Gemma
Git Séance integrates Gemma through Ollama for explanations grounded in retrieved Git-history evidence and for creative deleted-file epitaphs.
The model is part of the repository-investigation workflow, not simply an unrelated chatbot attached to the interface.
Best Use of Render
Top comments (0)