jAIlbreakO.S

jAIlbreakO.S began as the Antigravity Orchestra — a Python-based system where different scripts and functions each played a specialised role, strictly coordinated to handle tasks across development, content, and automation. It worked, but it had no structured routing and no way to chain agents intelligently. The instruments played independently; they couldn’t truly listen to each other. The reimagining asked one question: what if every instrument had a brain, a name, a criminal backstory — and they all lived in a fridge?

The pivotal choice was rebuilding the entire system on LangGraph, which gave the project three things the original lacked: stateful graph execution, checkpointing, and typed message passing between nodes. Each agent became a LangGraph node with a typed BaseState TypedDict, tenacity-backed retry logic, and a persona defined in personas/config.py. The supervisor graph (~1,800 lines) is the central nervous system. Every request hits a two-stage intent extractor — first keyword matching via ROUTING_RULES, then an LLM fallback using Claude Haiku for cost-safe disambiguation — before dispatching to the right agent.

What started as 67 agents grew to 108 across 15 specialised categories: Core Dev (8) — Architecture, code review, agent building Infrastructure (9) — DevOps, deployment, automation, MCP Data & Research (11) — Supabase intelligence, RAG, GitHub analysis Security & Quality (7) — Auditing, fact-checking, eval judging Finance (9) — Analytics, risk, investment, pricing Marketing & Content (12) — SEO, copy, social, email, video briefs Betting Ecosystem (7) — Football, darts, F1, horses, roulette Meditation Pipeline (7) — Script to YouTube, fully automated Audio & Video (6) — Whisper, Chatterbox TTS, FFmpeg, YouTube They argue through graph edges until they agree on a solution.

The most architecturally significant addition was the Memory Spine — a pgvector layer backed by Supabase, giving every agent persistent, searchable memory across sessions. Memories are embedded using text-embedding-3-small (1536 dimensions), stored with HNSW indexes for fast ANN retrieval, and scoped per-agent via hardened row-level security. A decay engine reduces memory strength 10% per cycle and archives anything that drops below 0.15 after 14 days. Five memory types: episodic, semantic, procedural, reflective, and shared (broadcast channel readable by all 108 agents). The result is a system that genuinely learns.

RALPH (Routing And Learning Pattern Handler) is an autonomous background process that continuously routes tasks, observes outcomes, and adjusts its own routing weights based on results. It’s the system learning how to dispatch itself better — a feedback loop layered on top of the supervisor graph.

End-to-end meditation video pipeline — script to YouTube with zero human involvement: 1. Theme Factory — generates theme and tone 2. Scriptwriter — writes the full guided meditation script 3. Narrator — voices it via Chatterbox TTS or ElevenLabs 4. Music Generation — scores ambient background audio 5. Visual Designer — selects and assembles background video 6. Video Assembler — FFmpeg rendering 7. YouTube Upload — titles, tags, description, publish A single CLI command triggers the full chain.

455 passing tests across routing integrity, agent load, eval gate behaviour, memory store, and pipeline execution. An LLM-as-judge eval gate scores outputs 1–10 on relevance, completeness, and clarity, failing closed if quality drops below threshold. The full stack runs containerised via Docker, served through FastAPI.