VORTEX-OS – A Hierarchical, Self-Healing Multi-Agent Engine for Complex Studio Production (With Deep-Sleep HITL Gates & Local Audio Pipeline)
​
Hey everyone,
I built a multi-agent framework to solve a problem I keep hitting with standard agent chains: context drift, infinite loops, and things just going off the rails on complex tasks.
It’s called **VORTEX-OS**, and it’s essentially a native orchestration engine designed to treat AI agents like a strict, closed-loop crew instead of a chaotic chatbot swarm.
To keep the AI from hallucinating or forgetting the plan halfway through, I built a rigid **4-Tier Chain of Command** (General Manager → Store Supervisor → Shift Supervisor → The Crew). This completely isolates context at every level so workers only see exactly what they need to build.
You can try out the skill.sh file with minimax M3 at this Minimax link
Here is a look at what the engine actually does under the hood, focusing on two main domains I've been refining:
### 1️⃣ Advanced Narrative & Worldbuilding
You give the engine a single natural-language master objective (like designing an entire interactive web scene), and it decomposes it across specialized domains:
* **Visual Novel Assets:** It generates individual scenes, character dialogue, ambient audio loops, and interactive HTML pages with functional choice mechanics.
* **Branching CYOA Systems:** It builds multi-paragraph branching stories, custom chapter selectors, and matching music prompts out of the box.
* **The Continuity Engine:** This is the best part. Every piece of narrative output is automatically audited against a strict universe canon and era constraints (e.g., if it's a 1994 setting, mentioning a smartphone triggers a hard fail). If an agent breaks canon, a **Self-Healing Prompt Optimizer** intercepts it, rewrites the prompt on the fly to forbid that specific failure mode, and re-dispatches the worker.
### 3️⃣ Local Sound Design, Foley & Signal Processing
I got tired of relying on fragile external web APIs for media processing, so VORTEX-OS handles complex asset rendering and processing locally inside an ephemeral sandbox:
* **Batch Sample Chopping:** It spins up Python and Bash workers to automate ffmpeg workflows, slicing up raw field recordings into crisp percussive hits.
* **Aesthetic Audio Degradation:** It dynamically chains ffmpeg filters (lowpass, wow, flutter, distortion) to emulate hardware constraints, meaning it can take clean voice tracks and instantly give them a degraded, warbly VHS tape or cassette texture.
* **Seamless Loops:** It automates crossfading math behind the scenes to create perfect, infinite atmospheric background drones without clipping or popping.
### 🛡️ Preventing Runaway Agents: The Deep-Sleep HITL Gate
To make sure a worker doesn't accidentally burn through millions of tokens or overwrite important files when a prompt drifts, I built a **Deep-Sleep Human-in-the-Loop gate**.
Before any high-stakes file write, the engine dumps its entire memory and execution state to disk, throws a PENDING_HUMAN code, and goes completely to sleep. It stays suspended until you review the logs and manually run ./skill.sh --hitl-approve. Everything runs locally via a standard CLI entry point (skill.sh), enforcing strict read/write boundaries across 8 contract invariants.
I’d love to get your thoughts on this hierarchical approach to agent governance. How are you guys tackling prompt self-healing and maintaining strict canon alignment in long-form creative projects?
The full file architecture, installation checks, and complete command reference are wrapped up in the master document