Marko Sabec

QuantForge separates the engine from the outside world with four traits: MarketDataSource, TradingVenue, CandleStore, and RunJournalStore. Binance Spot and SQLite are adapters behind them. Strategies sit behind a determinism contract: a pure function of the bar sequence, no clock, no randomness, no I/O, and decisions only on closed candles.

Because strategies are deterministic, a backtest and a live run over the same bars make the same decisions. A restart replays recent bars to warm indicators up, and backtests fill at the next bar's open. The StrategyContext is plain owned data with no borrows into the engine, so the boundary can later cross an FFI line into Python.

The live engine holds its venue as Option<&dyn TradingVenue>, and a dry run passes None. There is no flag to check before sending: there is nothing to send to.

Going live takes --mode live, credentials, and either --yes or a typed confirmation. A non-interactive run without --yes prints a preview of what it would do and exits cleanly, and the production host is labelled as such in the output.

After a fill, the journal write can fail. QuantForge defers that error until the position change is persisted: an executed order must exist in local state, otherwise a restart sees a flat position and doubles the exposure. Unsellable remnants below the venue minimum are written off so a run never wedges on dust.

Quantities are Decimal end to end and rounded down to the symbol's step size and lot rules before an order leaves the process. Time arithmetic uses checked operations, so an overflow is an error rather than a silently wrapped timestamp.

The e2e harness spawns the real binary with a cleared environment and an allowlist, so a sourced .env holding live keys can never leak into a test. The offline tier runs against a scriptable mock Binance server, and the testnet tier only runs when credentials are present.

CLI output is grep-friendly key: value lines on stdout with logs on stderr, and the absence of a run_id line means the bot did not run. The command tree and help output are snapshot-locked, so every CLI change shows up as a reviewed diff.