QuantClaw

The clockwork heart of the floor. Fires cycle triggers on a configurable cron, watches for market opens and closes, and routes the "wake up" signal to whichever agents the current phase needs. Doesn't think — just ticks.

Subscribes to every event on the bus and pattern-matches against a rules table — agent failure streaks, drawdown warnings, regime changes. Fires alerts routed to Telegram / Discord / Slack by urgency. Never planned, always listening.

Given a goal, proposes factor families, alternative datasets, and model types to explore. Calls the web_search tool to pull papers and news, then hands concrete suggestions to the ingestor and miner.

Pulls OHLCV and fundamentals from twenty data providers through a shared range-aware parquet cache. Only fetches the missing slice of any request — no duplicate downloads, and any range older than seven days is frozen immutable.

Generates candidate factor expressions with an LLM, evaluates each in the sandbox, then mutates the top performers across several generations. Every expression runs in a subprocess with AST validation and hard resource limits.

Always the last step of any plan. Reads every upstream result in the pipeline and writes a human-readable summary into the chat narrative. No actions of its own — pure synthesis.

Trains models on the discovered factors — gradient boosting, xgboost, lightgbm, lstm, transformer, and more. Emits a concrete Strategy class with signals() and allocate() implemented, saved to data/strategies/ for the validator and executor to pick up.

Replays a strategy against historical data with realistic costs, then re-runs it on a held-out window the trainer never saw. Combines what used to be two separate agents (backtester + evaluator) into one with two task modes: `backtest` for in-sample replay only, `validate` for the full flight check with overfit verdict.

In paper phase, loads each active strategy, computes target weights, aggregates them across deployments, and submits rebalance orders through the paper broker. In live mode, every order batch goes through compliance and risk gates first.

Runs continuously. Tracks portfolio-level drawdown, per-position sizing, sector concentration, and leverage. Can veto the executor before any order reaches the broker if a metric crosses its configured limit.

Not an LLM — pure rule evaluation. Checks each order batch against jurisdiction rules, account restrictions, and sanctioned-instrument lists. Blocks violations before the broker ever sees them.

Only runs when something has failed. Given the failing step's error context, stack trace, and surrounding cycle results, proposes a root cause plus an adjustment for the next cycle. Purely reactive — never part of a happy path.