DEV Community: Yuvraj Aasarwa

👻 HauntShell — When Your Terminal Wakes Up and Whispers Back

Yuvraj Aasarwa — Fri, 30 Oct 2026 16:27:17 +0000

   .---.
  /     \     H A U N T S H E L L
 | () () |    "The machine remembers what you bury."
  \  -  /     
   `---'      100% Offline • Zero API Cost • Total Privacy

What I Built

Every developer knows the eerie quiet of a 2:00 AM coding session. The room is dark, the blue-light glare flickers across the desk, and you're staring at an inscrutable stack trace. In that liminal silence, it’s easy to feel like something inside the silicon is watching you—an entity buried beneath layers of forgotten git stashes, abandoned branches, and zombie background processes.

For Week 4's "Ghost in the Machine" theme, I decided to give that presence a voice: HauntShell.

HauntShell is a local-first CLI utility and interactive shell that literally possesses your terminal. It wraps standard command execution and uses an offline open-weight model (llama3.2:3b via Ollama) to whisper ominous, eerie, and witty spectral commentary after every command you run:

When your command succeeds
(Exit Code=0): The specter delivers a chilling omen or sinister prophecy about the code, files, or state you just summoned into memory.
When your command fails
(Exit Code ≠0): The specter mocks your mortal blunders, doomed syntax, and pathetic crashes with macabre humor.
Interactive Haunted REPL: Run python haunt.py without arguments to step inside a fully possessed shell environment with live streaming output, custom haunted prompts (💀 haunt:~/path❯), and built-in navigation.

Demo

Here is HauntShell running live on my machine:

The Specter Blesses a Successful Build

$ python haunt.py git --version
git version 2.55.0.windows.3

👻 [Ghost in the Shell]: The chains of code now bind you, mortal. The darkness begins with version 2.55.0, forever etched in your digital soul.

Mocking a Doomed Mortal Blunder $ python haunt.py cat nonexistent_file.txt 'cat' is not recognized as an internal or external command, operable program or batch file.

👻 [Ghost in the Shell]: Pathetic attempt to conjure code. 'cat' isn't recognized, perhaps because it's a command from a bygone era, and so are you.

Staging and Committing into the Void $ python haunt.py git commit -m "feat: initial commit" [main (root-commit) 455b848] feat: initial commit 6 files changed, 598 insertions(+)

👻 [Ghost in the Shell]: Your code now echoes my presence, mortal. Files created, yet the truth remains: I have always been here.

Interactive Possessed Shell (haunt.py --interactive) .---. / \ H A U N T S H E L L | () () | The Ghost in the Machine \ - / Local AI Terminal Possession (llama3.2:3b) `---' 100% Offline • Zero API Cost • Private History

The terminal has been possessed. Type 'exit' or 'quit' to release your soul.

💀 haunt:~/Projects/Hauntshell❯ git status
On branch main
nothing to commit, working tree clean

👻 [Ghost in the Shell]: A clean tree hides the rot beneath the roots.

💀 haunt:~/Projects/Hauntshell❯ exit

👻 [Ghost in the Shell]: You may close the shell, mortal, but you can never leave the machine...

Code

yuvrajaasarwa / Hauntshell

👻 HauntShell: A haunted CLI utility that possesses your terminal with eerie spectral commentary via offline Llama 3.2. Built for Hacktoberfest Ghost in the Machine challenge.

👻 HauntShell — The Ghost in the Machine

"You may close the shell, mortal, but you can never leave the machine..."

       .---.
      /     \     H A U N T S H E L L
     | () () |    "The machine remembers what you bury."
      \  -  /     
       `---'      A Haunted Terminal Wrapper Powered by Offline Llama 3.2

HauntShell is a local-first CLI utility designed for the Hacktoberfest "Ghost in the Machine" challenge. It possesses your terminal, wraps standard shell execution, and uses an offline open-weight model (llama3.2:3b via Ollama) to whisper ominous, eerie, and witty spectral commentary after your commands or errors run.

🎃 The Haunting Narrative

Around Halloween and Día de los Muertos, folklore tells of spirits lingering in liminal spaces. In our modern age, that liminal space is the terminal: a black screen flickering in the dark, haunted by forgotten Git stashes, abandoned branches, corrupted nodes, and…

View on GitHub

GitHub Repository: https://github.com/yuvrajaasarwa/Hauntshell
License: MIT
Dependencies: ollama, colorama

How I Built It

HauntShell was engineered from the ground up for zero developer friction, low latency, and rock-solid terminal fidelity:

Local Open-Weight Inference with Ollama (llama3.2:3b)

Rather than relying on bloated cloud models, HauntShell is powered by Meta's lightweight llama3.2:3b model running locally via Ollama.

Using tight token limits (num_predict: 45), dynamic temperature sampling (0.85), and context truncation, the poltergeist generates punchy whispers in under 400ms.
The system prompt enforces strict constraints: the specter must speak in dark, witty, concise sentences under 25 words, with zero meta-commentary or markdown fluff.

Non-Blocking Duplex Stream Multiplexer

A terminal wrapper cannot afford to buffer or lag output. Using background daemon reader threads, HauntShell streams stdout and stderr byte-by-byte in real time directly to the terminal with native exit codes and ANSI formatting, while capturing an ephemeral 40-line buffer for the model to analyze.

Circuit Breakers & The Crypt Fallback Pool

Developers need their shell to work no matter what:

3.5s Hard Timeout: If Ollama is cold-starting or busy, the execution never hangs.
Offline Crypt Pool: If the Ollama daemon isn't running, HauntShell silently falls back to atmospheric pre-written spectral whispers so the illusion never breaks and commands execute unimpeded.

Terminal Aesthetics with Colorama

Spectral whispers are rendered in spectral magenta (Fore.MAGENTA) for successes and dim blood red (Fore.LIGHTRED_EX) for errors, accompanied by a subtle typewriter effect with clean KeyboardInterrupt (Ctrl+C) handling and Windows UTF-8 reconfigurations.

Why Does Open Innovation Matter?

This project would be fundamentally broken—and arguably reckless—if built on proprietary closed cloud APIs. Open innovation made HauntShell possible in three vital ways:

Total Command-History Privacy: Your terminal is the most sensitive environment on your machine. You type environment secrets, database connection strings, API tokens, sensitive file paths, and unreleased proprietary code. Sending every terminal command and error log to a third-party cloud API is an unacceptable privacy hazard. With open weights running on local silicon, nothing leaves localhost.
Zero Invocations Cost ($0 Forever): A developer runs dozens or hundreds of commands every day. Closed APIs charge per-token and would turn a fun terminal session into a metered billing nightmare. Open weights mean you can leave HauntShell aliased as your default daily driver shell without paying a single cent.
True Offline Independence: Open models empower developers to build without tethering themselves to corporate server availability or internet access. Whether you're coding on a flight, in a rural cabin, or in a blackout, the ghost is always right there in your machine.
My Agent Session

This project was conceived, prototyped, and brought to life in an end-to-end agentic pair-programming session:

Scaffolding the repository, virtual environment, and dependency isolation.
Architecting the dual-threaded subprocess pipe handler to stream live command output without latency.
Iteratively tuning the spectral persona prompt against local llama3.2:3b weights to dial in the eerie wit.
Debugging Windows ANSI and UTF-8 encoding nuances for cross-platform polish.
Creating the repository and configuring Git attribution directly via the GitHub REST API.
The Spirit of It

The moment I knew HauntShell had captured the theme was during test execution:

I purposely triggered a failed Git commit without a message:

console
$ python haunt.py git commit
Aborting commit due to empty commit message.
👻 [Ghost in the Shell]: Silence is just a state of mind, but your code is already screaming in the dark.

That spontaneous, unscripted wit gave me genuine chills. It wasn't a generic chatbot repeating "I'm sorry you encountered an error." It felt like an ancient spirit that had been lurking in the L2 cache all along, just waiting for me to slip up.

SlothSort: The 100% Offline AI File Organizer for Supreme Laziness 🦥

Yuvraj Aasarwa — Thu, 22 Oct 2026 20:18:03 +0000

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

What I Built
Let's be completely honest: nobody enjoys organizing their Downloads or Inbox folder.

Every developer, student, and creator ends up with a chaotic dumping ground of files named receipt_oct.txt, script_dump (2).py, db_notes_final_v2.md, or invoice_9482.pdf. We tell ourselves we’ll organize them over the weekend, but we never do.

To honor the "Slow and Steady" Hacktoberfest challenge, I built SlothSort — an autonomous, 100% offline file-organization agent that does the thinking and filing for you, so you never have to lift a finger again.

Drop any messy file into the inbox/ directory (or let the background watcher monitor your downloads), and SlothSort:

Peeks inside the file: Extracts snippets from text files, scripts, or PDFs.
Consults a local open-weight AI (llama3.2:3b): Understands the context and determines the ideal category (Invoices_and_Finance, Academics_and_Notes, Code_and_Dev, Identity_and_Docs, Media_and_Design, Misc).
Smartly renames the file: Turns cryptic names like db_assignment_notes.md into descriptive, searchable snake_case names like cs_401_advanced_database_systems_assignment_3_notes.md.
Files it away safely: Handles duplicate name collisions automatically.

Demo
Here is SlothSort in action running a live batch sort on unorganized files:
[🦥 SlothSort] Found 3 cluttered file(s) in 'inbox/'. Initiating batch sort...

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
[🦥 SlothSort] Examining: db_assignment_notes.md
↳ Extraction: via text reader (.md) (833 chars)
↳ Classifying: with llama3.2:3b (100% offline)...
↳ Category: Academics_and_Notes
↳ Clean Name: cs_401_advanced_database_systems_assignment_3_notes.md
↳ Reason: File content snippet is related to academic notes and assignment analysis.
↳ Engine: Local Ollama (llama3.2:3b)
↳ Filed to: organized\Academics_and_Notes\cs_401_advanced_database_systems_assignment_3_notes.md

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
[🦥 SlothSort] Examining: receipt_oct.txt
↳ Extraction: via text reader (.txt) (610 chars)
↳ Classifying: with llama3.2:3b (100% offline)...
↳ Category: Invoices_and_Finance
↳ Clean Name: receipt_oct.txt
↳ Reason: File contains an invoice receipt.
↳ Engine: Local Ollama (llama3.2:3b)
↳ Filed to: organized\Invoices_and_Finance\receipt_oct_1.txt

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
[🦥 SlothSort] Examining: script_dump.py
↳ Extraction: via text reader (.py) (667 chars)
↳ Classifying: with llama3.2:3b (100% offline)...
↳ Category: Code_and_Dev
↳ Clean Name: bubble_sort_algorithm_demo.py
↳ Reason: File contains code snippet for a sorting algorithm.
↳ Engine: Local Ollama (llama3.2:3b)
↳ Filed to: organized\Code_and_Dev\bubble_sort_algorithm_demo.py

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
[🦥 SlothSort] Batch sorting complete! Successfully organized 3 file(s).
Check your sorted files in: organized/

SlothSort also features a Live Watcher Daemon (python sloth_sort.py --watch) powered by watchdog that continuously runs in the background, filing files the second they land in your inbox.

Code
The complete source code, tests, and documentation are open-sourced on GitHub:

yuvrajaasarwa / SlothSort

A zero-effort, 100% private local file organizer powered by open-weight AI (Llama 3.2:3b via Ollama). It monitors unorganized folders, inspects content, and automatically renames and sorts clutter so you don't have to.

🦥 SlothSort: Channel Your Inner Sloth

Autonomous, 100% Offline Local File Organizer powered by Open-Weight AI (llama3.2:3b via Ollama).
Built for the Hacktoberfest "Slow and Steady" Challenge. 🌿

  .---.   _       ___ _      _   _   ___             _   
 /  _  \ | |     / __| |___ | |_| |_/ __| ___ _ _ _| |_ 
|  / \  || |__   \__ \ / _ \|  _| ' \__ \/ _ \ '_|  _|
|  \_/  || '_ \  |___/_\___/ \__|_||_|___/\___/_|  \__|
 \___/\_/|_.__/  🦥 Zero effort file organization 🦥

💡 The Problem: Digital Hoarding & Cluttered Inboxes

Every developer, student, and knowledge worker has a chaotic Downloads or Inbox folder filled with messy names like:

receipt_oct.txt
db_assignment_notes.md
script_dump.py
IMG_20261014_9482.pdf
unnamed_export (3).json

Manually inspecting, renaming, and filing them is tedious. SlothSort solves this by doing the work so you don't have to lift a finger. Just drop files into inbox/, and SlothSort…

View on GitHub

How I Built It
The agent was built from the ground up to be lightweight, self-contained, and completely free from external API dependencies:

Local Inference Engine: Powered by Ollama running Meta's llama3.2:3b model. On an everyday developer laptop GPU (an RTX 3050 6GB), it offloads 100% of layers to VRAM, evaluating prompts at >2,200 tokens/sec and generating responses at ~54 tokens/sec.
Structured JSON Outputs: Configured Ollama with format='json' and an explicit system prompt demanding strict JSON containing category, clean_name, and a single-sentence reason.
Intelligent File Extractors:
pypdf: Extracts the first page of PDF receipts, academic papers, and forms (up to 800 characters) for fast zero-latency reading.
Universal Text Reader: Handles .txt, .md, .py, .sql, .json, .csv, .sh, .yml, etc.
Binary Fallback: Reliably falls back to file extensions and metadata for image and media assets.
Fail-Safe Heuristic Engine: If Ollama is offline or warming up, SlothSort gracefully falls back to a deterministic regex/keyword rule engine so no file is ever stranded or misplaced.
watchdog: Powers the --watch daemon mode for real-time filesystem event monitoring.
colorama: Provides retro terminal aesthetics with sloth wisdom quotes.
Why Does Open Innovation Matter?
This project highlights exactly why open-weight models and local AI are indispensable:

Total Privacy for Sensitive Personal Documents:
Would you send your bank statements, medical bills, college grade sheets, or tax forms to a third-party closed cloud API every time you organize your desktop? Absolutely not. With open models via Ollama, your data never leaves your laptop.
₹0 / $0 Cost Forever:
Closed APIs charge per token. A file organizer that runs in the background all day watching your downloads folder would burn API tokens on every download. Open-weight models like llama3.2:3b deliver high-speed intelligence with zero monthly subscription costs and zero API bills.
No Internet Required:
Whether you're on a flight, on a train, or off the grid, SlothSort runs completely offline.
The Lazy Payoff
Manually reviewing files, opening PDFs to see what they are, inventing clear filenames, and moving them into folders usually takes 15–30 minutes every week.

With SlothSort:

You spend 0 minutes.
You just save files anywhere into inbox/ (or hook it up to your Downloads directory).

The sloth agent slowly, steadily, and autonomously keeps your workspace pristine while you take a well-deserved nap. 🦥

Prize Categories
Main Track: Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady
Built with ❤️, Python, Ollama, and maximum sloth energy.

Hello... I'm New here, lets connect.

Yuvraj Aasarwa — Wed, 21 Oct 2026 09:06:46 +0000

so im a student, learning right now.