DEV Community: Ganesh Patel

Git Séance — Resurrect the Lost Intent in Your Git History

Ganesh Patel — Wed, 28 Oct 2026 12:54:51 +0000

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.

Youtube demo - https://youtu.be/q1ym6HuCpBQ

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

solo dev

I stopped reviewing PRs after 8 PM: How an open Gemma agent audits diffs while I sleep

Ganesh Patel — Mon, 26 Oct 2026 06:58:21 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady

What I Built

It was 8:45 PM on a Wednesday. My eyes hurt from staring at terminal logs all day, my brain was completely fried, and my dinner was getting cold.

Then came the Slack notification:
"Hey! Just opened a quick PR for user auth and DB pooling. Only 380 lines. Mind doing a fast review before tomorrow's release?"

We've all been there. You know you should check every SQL string interpolation, verify async error boundaries, and hunt down unhandled null pointer exceptions. But your inner sloth is screaming: "I just want to lie on the couch and watch YouTube."

In previous years, being lazy meant either rubber-stamping the PR with a blind "LGTM, ship it" (and getting woken up by PagerDuty at 3 AM when production explodes), or spending an exhausting hour nitpicking code when you'd rather be resting.

So for Hacktoberfest Week 3 (Slow & Steady / International Sloth Day), I channeled my inner sloth and built SlothMerge: The Ghost PR Reviewer.

Code

Ganeshpatel404 / slothmerge

Local open-source AI agent for auditing PR diffs using Gemma 2 and Sentry Agent Tracing

SlothMerge: The Ghost PR Reviewer

"Build something with open-weight models that rewards laziness."
SlothMerge is an open-source AI agent powered by Google Gemma 2 and instrumented with Sentry Agent Tracing. It intercepts git diffs, flags security and logic flaws, computes financial savings versus proprietary APIs, and outputs an executive 3-bullet morning digest and pre-formatted GitHub PR review comment.

01. Why Open Innovation Matters Here

Zero Source Code Leaks: Proprietary diffs, database schemas, and business logic stay on your laptop or inside your private VPC. No source code ever touches external AI providers.
Zero Recurring API Costs: Instead of paying $0.03 to $0.15 per multi-file PR review on closed APIs, SlothMerge runs on open-weight Google Gemma 2 models (gemma2:2b or gemma2:9b) costing $0.00.
Deterministic & Local: Works completely offline on an airplane or when network connectivity is down.

02. Architecture & Pipeline

[ Git Diff / PR

…

View on GitHub

How I Built It

SlothMerge is built around three core pillars:

Open-Weight Foundation: Google Gemma 2 I used Gemma 2 (gemma2:2b and gemma2:9b). Gemma 2's sliding-window attention and strong code-comprehension benchmarks make it ideal for parsing compact unified diffs. The prompt uses Gemma's structured turn syntax:

text

user
You are SlothMerge...

model
SlothMerge supports running directly on local hardware via Ollama (http://localhost:11434) as well as an optimized fallback pipeline when running in containerized or offline environments.

Sentry Agent Tracing (sentry-sdk) Code review agents can be notoriously opaque. SlothMerge uses Sentry Agent Tracing to track every operation into distinct spans:

diff.parse: Extracting changed files, line numbers, and hunks.
ai.inference: Prompt construction, model latency, and token throughput.
output.format: Markdown serialization and clipboard formatting.

Why Does Open Innovation Matter?

Why Does Open Innovation Matter?
Why not just hook up a commercial closed API like GPT-4o or Claude 3.5 Sonnet to a GitHub webhook?

Here is why open innovation made SlothMerge possible in ways a closed API could not:

Zero Code Leakage (Total IP Protection)
Your company's unreleased source code, proprietary algorithms, database schemas, and internal API endpoints should never be sent to a third-party closed API server. With Gemma running locally or in your private VPC, your diff never leaves your machine.
Zero Recurring API Costs
A busy engineering team might open 40–80 PRs a day. At ~2,000 to 8,000 tokens per comprehensive diff review, closed APIs quickly accumulate hundreds of dollars a month in recurring costs. Running open-weight Gemma 2 costs exactly $0.00.
Reliable Offline Operation
SlothMerge works on an airplane, on a train, or during an internet outage. When the Wi-Fi drops, your code review pipeline doesn't stop

The Lazy Payoff

The Lazy Payoff
Here is the real math behind the sloth approach:

Average PR review time for a human engineer: 25–40 minutes.
SlothMerge audit latency: ~350 milliseconds.
Energy expended by human: 0 calories.
This week alone, SlothMerge audited 3 candidate PRs:

Caught an accidental personal access token commit that would have required revoking credentials.
Flagged an unescaped SQL query that would have introduced a SQL injection vulnerability into user authentication.
Approved a routine string helper refactor instantly, freeing the author to merge without waiting 18 hours for someone to get around to reading their code.

Prize Categories

Sniff Quest-AI-adventures-for-dog-walks

Ganesh Patel — Mon, 12 Oct 2026 08:07:38 +0000

SniffQuest: AI Canine Olfactory Quests That Get You and Your Dog Touching Grass 🐾🌿

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

What I Built

Dogs experience up to 90% of their world through scent. A dog's sense of smell is roughly 10,000 to 100,000 times more acute than ours, and just 20 minutes of mental scent exploration ("sniffari") tires and enriches a dog more than an hour of mindless physical pavement trotting.

Yet modern dog walks have a common problem: the human is staring down at a glowing phone screen while rushing the dog along on a tight leash.

SniffQuest is an open-source, outdoor enrichment web application designed to break this cycle. It turns everyday walks into captivating sensory quests, olfactory detective missions, and screen-free outdoor adventures tailored to your pup's breed, energy level, and surroundings.

Key Highlights:

🌲 Canine Adventure Planner: Generates personalized sensory waypoints (mossy tree bark, damp pine needles, sunny clover patches, fallen logs) based on your dog's profile and terrain.
🚶 Screen-Free "Walk Mode": Designed to put your phone in your pocket! A clear, high-contrast outdoor HUD lets you track sniffing waypoints, log olfactory milestones, and follow your dog's nose without staring at a map.
🏅 Enrichment Badges & Local Journal: Celebrates real-world exploration by awarding local achievement badges (e.g., Pinecone Sleuth, Wildflower Trailblazer) with 100% private, client-side history storage.
🛡️ Canine Safety Guidelines: Built-in veterinarian-backed safety guidance, including 7-second asphalt temperature tests, loose-leash pacing, and toxic flora awareness.

Demo

Check out the project running locally or clone the repository to run it with zero dependencies:

GitHub Repository: https://github.com/Ganeshpatel404/AI-adventures-for-dog-walks
Live Local Setup:


bash
  git clone https://github.com/Ganeshpatel404/AI-adventures-for-dog-walks.git
  cd AI-adventures-for-dog-walks
  npm install
  npm run server & npm run dev