DEV Community

Wellington Filipe
Wellington Filipe

Posted on

A ghost in my shell, powered by a local Gemma

#devchallenge #hf26challenge #opensource #ai

This is a submission for the Hacktoberfest Open-Source AI Challenge: Week 4 — theme: Ghost in the Machine.

What I built

Haunted Terminal: a ghost that lives in your shell. Every few commands (1 in 4 by default), it whispers one eerie sentence about what you just ran and whether it failed.

$ rm -rf node_modules
👻 A clean sweep… but something vital vanished with those modules, didn't it?
Enter fullscreen mode Exit fullscreen mode

It is about 80 lines: a Bash hook (haunt.sh) on PROMPT_COMMAND and a stdlib-only Python script (whisper.py) that calls Ollama's /api/generate. No dependencies, no API key, no account.

Code: https://github.com/Tinhomagri/haunted-terminal

Why open-weight matters here

  • Your shell history is private. Commands carry hostnames, paths, tokens pasted by mistake. A ghost that reads them must not be a cloud API. Here the prompt never leaves 127.0.0.1.
  • It works with the lights off. No internet, no outage, no rate limit. If Ollama is down the ghost just stays silent (except Exception: pass) and the prompt is never blocked, because the call runs in a background subshell.
  • I can swap the model with one env var (HAUNT_MODEL). That let me compare two sizes honestly instead of trusting a spec sheet.
  • Costs nothing per command. A hook that fires on every few keystrokes would be absurd to bill per token.

Gemma 3 4B vs 12B (measured, not guessed)

Same prompt, 5 commands, 16 cores, no GPU, 14 GB RAM (warm = model already loaded):

Model First call Warm median Warm max
gemma3:4b 9.5 s 2.2 s 2.3 s
gemma3:12b 5.5 s (pre-warmed) 8.2 s 9.3 s

Two things I learned:

  1. 12B writes slightly better lines, but 4B is the right default. The ghost runs in the background, but a whisper that shows up 9 s after the command feels disconnected; 2 s feels like it is watching you.
  2. Cold start can silence the ghost. My first 12B run timed out: loading ~8 GB of weights exceeded the 20 s request timeout, and the "never explain yourself" error handling hid it. When it loaded first, the same call took 5.5 s. If you use a bigger model, keep it resident (OLLAMA_KEEP_ALIVE) or raise the timeout.

Sample 4B lines:

  • git push --force → The past resists your forceful rewrite, forever staining its digital echoes.
  • sudo apt update → The system whispers of updates, but something else is being quietly installed.

Try it 

ollama pull gemma3:4b
git clone https://github.com/Tinhomagri/haunted-terminal && cd haunted-terminal
echo "source $PWD/haunt.sh" >> ~/.bashrc
Enter fullscreen mode Exit fullscreen mode

Env: HAUNT_ODDS, HAUNT_MODEL, HAUNT_HOST.

Limits

Small models sometimes repeat "echoes" and "something unseen". Bash only (zsh needs precmd). Latencies above are CPU-only on one machine.

Top comments (0)