Phantom Shell — a haunted terminal possessed by a local open-weight model

Ri Piao — Wed, 28 Oct 2026 06:41:51 +0000

Phantom Shell 👻 — Ghost in the Machine (Hacktoberfest 2026 Week 4)

Challenge: #devchallenge #hf26challenge — Week 4 Ghost in the Machine

Author: Ri Piao · X @pr742674

What I built

Phantom Shell is a CLI that lets an open-weight model possess your terminal.

You run a normal shell command. The phantom — a local model via Ollama — whispers a short, eerie commentary about what you just did. You can also chat with it, or drop into possess mode: an interactive haunted REPL.

If Ollama is not installed, a built-in mock ghost still demos the full UX so judges (and you) can try it offline in seconds.

Theme take: the “ghost” should live in the machine — open weights, local inference, no cloud API key. Privacy is the haunting.

Demo

git clone https://github.com/pr7426/ghost-in-the-machine.git
cd ghost-in-the-machine
python3 -m phantom_shell demo
# or:
./scripts/demo.sh

Interactive:

python3 -m phantom_shell possess
# phantom> echo boo
# phantom> .chat Who are you?
# phantom> .quit

With a real local model:

ollama pull llama3.2
python3 -m phantom_shell possess --model llama3.2

Sample transcript (checked into the repo): demo/sample_session.txt

Repository

🔗 Repo: https://github.com/pr7426/ghost-in-the-machine

License: MIT · Language: Python 3.10+ · Deps: stdlib only for core

Open-source AI used

Piece	Detail
Serving	Ollama (`/api/chat`)
Model	`llama3.2` by default (any Ollama open model works: `phi3`, `qwen2.5`, `gemma2`, …)
Offline	Mock ghost when Ollama is down
Optional path	llama.cpp / GGUF via Ollama or a local server (`--host`)

No closed SaaS LLM is required for the core experience.

Why open innovation matters (here)

Offline & private — command context and chat stay on-device; great for a personal shell.
Swappable weights — one flag changes the soul of the ghost; community models welcome.
Hackable — tiny stdlib client, MIT, easy to fork into prompts, tools, or a full agentic shell.
Inclusive demos — mock mode means anyone can run the challenge submission without a GPU.

Open weights turn “AI feature” into something you can own, inspect, and remix — which is the point of a Ghost that lives in your machine.

Thanks Hacktoberfest + DEV for Week 4 — may your shells stay haunted. 🎃

HermesSec: Turning Open-Weight Gemma & AST Triage into a $149 Freelance Security Audit Side Hustle

Ri Piao — Wed, 14 Oct 2026 18:21:12 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle.

What I Built

Indie SaaS founders, vibe-coders, and small dev agencies ship payment and authentication code at breakneck speed. They cannot afford a traditional $15,000 penetration test before launching, yet a single SQL injection (CWE-89), hardcoded production secret (CWE-798), or shell=True command execution (CWE-78) can wipe out their business overnight.

On the flip side, freelance security practitioners want to offer affordable $149 Express Pre-Launch Code Audits—but manually reviewing snippets, mapping CWE identifiers, and writing copy-pasteable remediation patches takes hours of unbilled effort.

HermesSec bridges this gap. Inspired by the Hermes / OpenClaw multi-agent task decomposition workflow, it combines deterministic Python AST (Abstract Syntax Tree) pre-triage with Google Gemma open-weight models, Backboard's stateful thread memory (memory="Auto"), and Render cloud hosting to generate line-accurate, client-ready security audit reports in seconds.

Demo

Live App on Render: https://hermes-sec-audit.onrender.com/
Try it live: Click Sample 1 (SQLi + Secret) or Sample 2 (RCE / Command Inj), or paste your own custom Python code—the AST engine + Gemma pipeline will dynamically detect the exact line numbers and generate a tailored patch!

Code

GitHub Repository: https://github.com/pr7426/hermes-sec-audit/tree/main
Includes app.py (FastAPI + AST Triage + Backboard /api/threads/messages + Local Ollama /api/chat), Dockerfile, and render.yaml Blueprint.

How I Built It

HermesSec uses a three-stage sub-agent architecture designed for zero-hallucination reliability:

Sub-Agent 1 — Deterministic Python AST & Regex Triage: Parses incoming source code into an Abstract Syntax Tree (ast.walk) to pinpoint exact line numbers for dangerous sinks (eval, exec, pickle.loads, subprocess with shell=True, raw SQL string interpolation, and hardcoded secrets).
Sub-Agent 2 — Google Gemma Remediation Engine: Feeds the structured AST findings into open-weight gemma-2-9b-it (via Backboard or local Ollama /api/chat) to synthesize drop-in secure refactoring code and executive risk summaries.
Sub-Agent 3 — Backboard Stateful Memory & Model Benchmark: Uses Backboard's POST https://app.backboard.io/api/threads/messages endpoint with X-API-Key and memory="Auto" to retain recurring CWE context across audit threads and compare google/gemma-2-9b-it against qwen-2.5-coder-7b-instruct side-by-side.

Why Does Open Innovation Matter?

A freelance code auditing business cannot be built on closed commercial AI APIs for three practical reasons:

Client NDA & Source Code Privacy: Paying clients will never allow proprietary authentication or checkout code to be sent to closed-source frontier providers that retain prompts for training. Because HermesSec supports open-weight Gemma via local Ollama (http://localhost:11434/api/chat) or controlled endpoints, sensitive client code stays completely private.
No False-Positive Safety Refusals on Security Code: Closed commercial APIs frequently refuse to analyze legitimate SQL injection queries, XSS sinks, or command-injection snippets due to overly aggressive refusal guardrails. Open-weight models like Gemma let security engineers inspect and patch real vulnerabilities without arbitrary blocks.
99% Gross Margin Unit Economics: Running a 9B open-weight model combined with local AST pre-filtering costs less than $0.0003 per audit, making a fixed-price $149 micro-audit highly profitable from day one.

My First User

I put the live Render deployment in front of a fellow developer and content creator who launches web apps and social e-commerce tools.

How it went: They pasted a real backend utility script into HermesSec. Within 2 seconds, the tool flagged a hardcoded token and an unvalidated query line by exact line number, outputting the parameterized replacement code.
Would they pay?: They said paying a flat $99–$149 for a verified pre-launch security sign-off before opening public access is an easy "yes" compared to enterprise consulting rates.
What they asked for next: Automatic GitHub Pull Request webhook integration and downloadable PDF certificates they can show to their customers.

Prize Categories

Best Use of Render: Deployed the containerized FastAPI + multi-agent runtime on Render using a native Dockerfile and render.yaml Blueprint.
Best Use of Gemma: Used Google's open-weight Gemma (gemma-2-9b-it / gemma2:9b) as the primary vulnerability explanation and secure patch synthesis model.
Best Use of Backboard: Integrated Backboard's POST /api/threads/messages API with X-API-Key and memory="Auto" to maintain stateful audit memory and compare open-weight models (gemma-2-9b-it vs qwen-2.5-coder) through a single key.

DEV Community: Ri Piao

Phantom Shell — a haunted terminal possessed by a local open-weight model

Phantom Shell 👻 — Ghost in the Machine (Hacktoberfest 2026 Week 4)

What I built

Demo

Repository

Open-source AI used

Why open innovation matters (here)

HermesSec: Turning Open-Weight Gemma & AST Triage into a $149 Freelance Security Audit Side Hustle

What I Built

Demo

Code

How I Built It

Why Does Open Innovation Matter?

My First User

Prize Categories