mirofish-offline-simulation — Claude Code skill: local swarm sim

What mirofish-offline-simulation does

The skill covers setting up MiroFish-Offline, an engine that takes a document — a press release, a policy draft, a financial report — and turns it into a graph of entities and relations using an LLM. From that graph it generates hundreds of agents, each with its own personality and memory, and lets them post, argue and change their minds on simulated platforms. A ReportAgent then interviews a focus group, searches the graph and writes up the analysis. You can also chat with any single agent, and its memory sticks.

Under the hood: a Flask backend on Python 3.11, a Vue 3 frontend on Node 18, Neo4j CE 5.15 over bolt, models served by Ollama through its OpenAI-compatible /v1 endpoint, and nomic-embed-text embeddings at 768 dimensions. Search is hybrid — 0.7 vector similarity plus 0.3 BM25. Two install paths exist: Docker Compose, or a manual build of Neo4j, Ollama, backend and frontend. Settings live in .env: LLM endpoint and key, Neo4j credentials, embedding model.

The main model qwen2.5:32b is roughly 20GB and wants 24GB of VRAM, so without that headroom pick qwen2.5:14b at 10GB. You can swap Ollama for any OpenAI-compatible provider by changing LLM_BASE_URL and the key. Simulation runs as a blocking call, so production use needs an async wrapper.