Arbos

Welcome! Arbos is simply a Ralph-loop combined with a Telegram bot.
That's all you need to do just about anything.

The Design

Arbos just loops a GOAL.md through a coding agent.

                                     ┌────── [GOAL.md] ────────┐
                                     ▼                         │
                ┌──────────┐     ┌───────┐                     │
                │ Telegram │◄───►│ Agent │─────────────────────┘
                └──────────┘     └───────┘

Requirements

• Telegram Bot token
• Chutes API key

Getting started

curl -fsSL https://raw.githubusercontent.com/unconst/Arbos/main/run.sh | bash

Usage

To run Arbos just set the /goal:

/goal

Use the below program to evolve a system S that discovers profitable trading strategies.

You are given:
C = { Hyperliquid capital, Coinglass derivatives data (funding, OI, liquidations, leverage), compute on Basilica/Targon/Lium }

Initial state (build first)
S₀ = online continuous adaptive trading system which:
    - uses online data for training
    - uses evolutionary model search (mutate/replace weak models)
    - uses strict walk-forward validation + online Sharpe filtering
    - uses horizon ensembles H = {1h,4h,8h,12h,24h}
    - uses consensus gating for signals
    - uses time-series foundation models

Run this loop continously
loop t = 1..∞
    S_t = design_or_modify(S_{t-1})  # implement current design S
    O_t = run(S_t)                   # run S: i.e. train, evaluate, trade
    P_t = measure(O_t)               # eval: Sharpe, PnL, drawdown, regime behavior
    Δ_t = reflect(S_t, P_t)          # find weaknesses in your design 
    S_{t+1} = improve(S_t, Δ_t)      # design a new design.
end

Then iterate.

---

MIT