Architecture

Four repositories and one for decisions:

arvo-engine	the daemon, the MCP server and every crate that decides anything
arvo-desktop	the window and its editor; links no engine crate, talks to the engine over the API
arvo-engine-api	the API: protos and the Rust and Python bindings generated from them
arvo-extension-api	the provider contract a plugin implements
arvo-adrs	decisions, research notes and plans, across all of them

The research loop

flowchart LR
    H["Hypothesis"] --> E["Experiment"] --> S["Simulation"] --> EV["Evaluation"]
    EV --> D["Evidence"] --> M["Research memory"] --> A["Agent"]
    A -.->|"asks the next question"| H

    style EV fill:#1f6feb,stroke:#1f6feb,color:#fff

NautilusTrader is the financial runtime underneath, linked in-process: orders, positions, accounts, execution, backtesting. Arvo owns what is unique to Arvo — out-of-sample selection, deflation, Inconclusive as a verdict, reconciliation, replayable findings, advice — and exactly one crate is permitted to name a Nautilus type. The differentiator is evaluation, not execution.

Standing rules

• The window knows nothing about how a study is run. It renders what the engine says. The build enforces it: the desktop depends on the two contract crates and nothing else of Arvo's.
• The engine keeps running when the window closes. Schedules run and sessions trade with nothing open; the window, the Python package, the MCP server and the command line are all clients of one engine per user.
• The AI never receives broker credentials. An agent reaches Arvo through a tool list that cannot trade.
• One risk policy, backtest and live.
• Bars are fetched to files and pinned by content hash. A study never reads a vendor live; a tick never becomes a bar.
• A provider is a process. Arvo starts it, hands it a token per spawn and a grant per call, and asks nothing of a service it did not claim. No third-party code runs in the window.
• Anything that makes a wrong answer look right outranks anything that adds capability.

Inside the engine

crates/ is the platform (data, research, risk, execution, the plugin host, the one Nautilus crate), integrations/ talk to vendors (Alpaca, Robinhood, Yahoo), app/ is the daemon and the MCP server. Traits live beside their implementer or their primary consumer, never in a shared types crate, and a provider trait is earned by removing a dependency or a panic — never by anticipating an implementation nobody asked for.