idolevi

That's fair, the PRD example is pretty cool but not the best - I have a few workflows that I use myself for doing bug bounties and deep research, but they are a bit much.

The current examples were mostly to show how you could set up different phases and how agents can work in parallel in different phases while still keeping some structure - basically any problem that you could define the logical phases in the way for solving it is going to work great with Hephaestus.

Soon I'll release a few more examples and a few tools that would make Hephaestus a development tool that learns about your repo as time goes and able to fix and add features for your code base

It's still a very first version honestly.. What did you find in the configuration hard? I'll try to streamline it so it'll be easier to run

Hey, This is just a very first version of Hephaestus that is only a month in a work to showcase the idea of an autonomous agent orchestration. As we speak I'm building on top of it a development tool that would learn your codebase as time goes on, which is hopefully going to result in faster development.

Although do note that Hephaestus is not only for development purposes, it's supposed to be an agnetic framework for building workflows - such as a research workflow, investment workflow and so on, based on the semi-structured phases principles.

Also - everything is forever going to be open source (including the future features and development tools), so look out :)

Hey everyone!

I've been working on Hephaestus - an open-source framework that changes how we think about AI agent workflows.

The Problem: Most agentic frameworks make you define every step upfront. But complex tasks don't work like that - you discover what needs to be done as you go.

The Solution: Semi-structured workflows. You define phases - the logical steps needed to solve a problem (like "Reconnaissance → Investigation → Validation" for pentesting). Then agents dynamically create tasks across these phases based on what they discover.

Example: During a pentest, a validation agent finds an IDOR vulnerability that exposes API keys. Instead of being stuck in validation, it spawns a new reconnaissance task: "Enumerate internal APIs using these keys." Another agent picks it up, discovers admin endpoints, chains discoveries together, and the workflow branches naturally.

Agents share discoveries through RAG-powered memory and coordinate via a Kanban board. A Guardian agent continuously tracks each agent's behavior and trajectory, steering them in real-time to stay focused on their tasks and prevent drift.

GitHub: https://github.com/Ido-Levi/Hephaestus Docs: https://ido-levi.github.io/Hephaestus/

Fair warning: This is a brand new framework I built alone, so expect rough edges and issues. The repo is a bit of a mess right now. If you find any problems, please report them - feedback is very welcome! And if you want to contribute, I'll be more than happy to review it!

I built a system that actually BLOCKS Claude from doing things you didn't ask for. Real-time violation detection with immediate intervention.

How it works: EVERY interaction gets analyzed - every "I'll help you with that", every tool call, every single message. Whether Claude is thinking, reading a file, or trying to edit code - it ALL goes through GPT-OSS (via Groq for near-instant analysis).

When you send a message, we extract your intent and add it to the session context. When Claude sends ANY message or tries ANY tool, we analyze if it aligns with your accumulated instructions. Every. Single. Time.

The magic happens in the pre-hook: Before Claude can execute ANY operation (Edit, Bash, Read, etc.), our system checks for violations. If Claude is doing something you didn't ask for - BLOCKED. The operation never executes. Claude gets a detailed explanation of what went wrong and why.

Examples it's caught:

- Asked Claude to analyze code → it tried to commit → BLOCKED

- Said "don't modify files" → Claude tried to edit → BLOCKED

- Requested a backend fix → Claude wandered into frontend → BLOCKED

- Asked to run tests → Claude said "I can't execute commands" → BLOCKED

The system uses "trajectory thinking" - GPT-OSS analyzes your ENTIRE conversation history to understand context. "Fix this" knows what "this" refers to from 20 messages ago. "Don't use external libraries" from the start stays enforced throughout. It understands multi-step workflows, so normal investigation isn't flagged.

Each Claude session maintains its own violation context - no cross-contamination between different conversations. Everything stored in SQLite, analyzed in real-time, blocked before damage.

...and it's also a Tamagotchi that lives in your statusline. Gets increasingly angry when Claude ignores instructions. Needs feeding every few hours. What started as a simple virtual pet somehow evolved into a full behavioral monitoring system. Feature creep at its finest.

Install: bun add -g claude-code-tamagotchi

Repo: https://github.com/Ido-Levi/claude-code-tamagotchi