Dragonfly

An institutional-grade, market-agnostic, polyglot quantitative trading system.

Six layers run as separate processes and communicate over a message bus. Each layer owns one stage of the path from market data to an executed trade. Money and state cross process boundaries only as fixed-size messages.

Layer	Name	Job	Language
L1	Ingestion	Turns venue data into normalized messages and keeps the book	C++
L2	Strategy	Turns market observations into trading intentions	C++ / Python
L3	Risk	Approves or rejects each intention	C++
L4	OMS	Records positions, cash, orders, and P&L	C++
L5	EMS	Sends orders to the venue and reports the outcomes	C++
L6	Observability	Shows what is happening and lets a human stop it	Python

Throughout these documents a layer is written as L<n> with its name in parentheses, such as L3 (risk) or L5 (EMS).

Documentation

Document	Read it for
docs/overview.md	The runtime picture: layers, processes, channels, and one tick end to end. Start here
docs/contracts.md	The message set: every field of every message, and the conventions they follow
docs/layers.md	What each layer does, what it must not do, and how it fails
docs/lifecycle.md	Three worked flows: a tick, an order, and a restart
docs/decisions.md	Why the system is built this way, one decision per entry
docs/operations.md	Build and run, the configuration files, and the phase plan

If you read one document, read docs/overview.md.

Status

Phase P0: the repository skeleton and toolchain. One smoke test passes. The message contracts are specified in docs/contracts.md but not yet implemented. That work is P1. The build sequence is in docs/operations.md, "The phase plan".

Toolchain

Concern	Choice
C++ compiler	Clang (Apple Clang locally)
Standard	C++20
Meta-build	CMake, driven by presets
Build tool	Ninja
C++ packages	vcpkg, manifest mode
Host tools	pkg-config, clang-tidy, clang-format
Python	3.11 or later, managed with uv
Testing	GoogleTest
Benchmarking	Google Benchmark

To build and run the tests, see docs/operations.md, "Build and run".