rift-engine-core 0.1.1

RIFT — Research / Iteration / Forecast / Trade. Quant trading infrastructure for Hyperliquid (CLI + research + MCP server).

RIFT

Quant Trading Infrastructure for Humans and AI. Research, validate, and deploy systematic trading strategies on Hyperliquid (HL) perps. Published by Nexstone.

⚠️ RIFT is software, not financial advice. Trading perpetual futures involves substantial risk of loss including the total loss of deposited capital. Past backtest performance does not predict future returns — markets drift, regimes change, models decay. RIFT ships with safety primitives (kill switches, drawdown limits, signed auth tokens, on-chain audit trail) but no software eliminates market risk. Never trade more than you can afford to lose, and never deploy a strategy you don't personally understand. The one strategy that ships (trend_follow) is an explicitly demo-only EMA crossover — see its docstring.

RIFT is built for two audiences in one tool:

• Quant-curious traders — tired of trading on gut. Want statistical backing for every decision but don't necessarily write Python. Use the workbench, CLI, or MCP-driven AI assistant.
• Top-tier quants — write Python directly against substrate primitives. Compose factor models, Kelly sizing, deflated Sharpe checks, alpha-decay measurement, and walk-forward validation without leaving the framework.

The same engine that powers the CLI also powers an MCP server, so Claude / other AI agents can drive the entire research-to-trade flow locally.

---

What RIFT does

• Backtest on real Hyperliquid candle, funding, and L2 order-book data
• Validate with purged k-fold CV, walk-forward, Monte Carlo, deflated Sharpe ratio
• Measure capacity (impact + ADV + L2-depth), alpha decay (half-life), cross-impact for baskets
• Promote / reject strategies through configurable gates (DSR, CV pass rate, drawdown, track record, capacity)
• Seal every result with content-addressed reproducibility manifests
• Execute live on HL with auth-token capability tiers, kill switches, audit trail
• Portfolio supervision: coordinate risk across multiple algo strategies
• One reference strategy (trend_follow) ships as a worked example that passes the framework's own promotion gates

---

Architecture

Three interfaces over five core engine layers, with auxiliary packages composed on top:

                                ┌──────────────────────────┐
   AI agents (Claude, etc.) ───>│  MCP server (rift serve) │ ─┐
                                └──────────────────────────┘  │
                                                              │
   Terminal users  ────────────>┌──────────────────────────┐  │
                                │  CLI (rift <command>)    │ ─┤  same Python
                                └──────────────────────────┘  │  engine + data
                                                              │
   Python power users ─────────>┌──────────────────────────┐  │
                                │  Python (import rift_*)  │ ─┘
                                └──────────────────────────┘
                                ┌──────────────────────────────────────┐
                                │ research — pipeline orchestrator     │
                                ├──────────────────────────────────────┤
                                │ engine — backtest / WF / MC / sweep  │
                                ├──────────────────────────────────────┤
                                │ trade — auth / propose / execute     │
                                ├──────────────────────────────────────┤
                                │ substrate — math primitives          │
                                │   (frictions, stats, validation,     │
                                │    decay, capacity, regime, etc.)    │
                                ├──────────────────────────────────────┤
                                │ data — HL S3 + REST + local cache    │
                                └──────────────────────────────────────┘

Three interfaces (CLI / Python / MCP), one engine. The diagram shows the five core engine layers (data → substrate → engine + trade → research); portfolio, strategies-sdk, api, and core compose on top. Whatever the CLI can do, the Python API and the MCP server can also do. Strategies are written once and run from any of the three.

---

Install

Requires:

• uv for Python (will auto-install Python 3.13 if your system doesn't have it)
• Node 20+ and pnpm for the CLI

# Clone
git clone <repo-url> rift
cd rift

# Python deps via uv (installs Python 3.13 + all workspace packages editable)
cd engine && uv sync && cd ..

# CLI deps via pnpm
pnpm install
cd packages/cli && pnpm build && cd ../..

# Configure secrets (AWS for S3 sync + optional HL wallet for live trading)
mkdir -p ~/.rift && cp .env.example ~/.rift/.env
# Edit ~/.rift/.env with your AWS access key + secret (free tier sufficient)

# Make the CLI binary callable as `rift` (one-time, optional but recommended)
# Pick one:
#   - Add the bin dir to PATH:
#       export PATH="$PWD/packages/cli/bin:$PATH"
#       (add to ~/.zshrc or ~/.bashrc to persist)
#   - Or symlink into ~/bin (or any dir already on your PATH):
#       ln -s "$PWD/packages/cli/bin/run.js" ~/bin/rift

You now have:

• rift (CLI) callable via packages/cli/bin/run.js — symlink or add to PATH (see above) for the bare rift shorthand
• rift-engine Python entry point at engine/.venv/bin/rift-engine
• 9 rift_* Python packages available via import (rift_core, rift_data, rift_substrate, rift_engine, rift_trade, rift_research, rift_portfolio, rift_strategies_sdk, rift_api)
• rift serve starts the MCP server (59 tools)

---

Quickstart — 60 seconds

# 1. Sync BTC 4h candles + funding from the Hyperliquid S3 archive
#    (default window: 2023-09-01 → today, ≈2 years). Requires AWS
#    credentials in ~/.rift/.env. One-time cost: ~$2 in S3 egress.
#    Skip this and the next step auto-fetches a smaller window (~6-10
#    months) from the HL REST API — same code runs, smaller sample.
rift sync --coins BTC --tf 4h

# 2. Run the full research pipeline on the bundled OSS reference strategy
rift research trend_follow --pair BTC --tf 4h

You will see the framework execute: backtest → walk-forward → Monte Carlo → multi-pair → feature importance → volatility forecast → health check → purged CV → alpha decay → capacity → promotion verdict → sealed reproducibility bundle.

With the sync above, trend_follow on BTC 4h returns +25.0% / Sharpe 0.71 / -6.88% max DD over ~2 years, passes 5 of 5 promotion gates (default config: EMA 50/200, 20% equity per trade). Without the sync, the auto-fetched ~6-10 month window gives smaller numbers — a recent fresh-install run showed roughly +5.7% / Sharpe 0.32 / fails 3 of 5 gates on the shorter sample. Same code, same verdict logic; the data window is what changes. The framework reports honestly on whichever you give it.

---

Layer-by-layer overview

Layer	Package	What it does
Core	rift_core	Shared types, config, schema, key management, NDJSON output protocol — the substrate the other Python packages share
Data	rift_data	HL S3 + REST sync, candle / funding / OI / L2-book loaders, local parquet cache
Substrate	rift_substrate	Math primitives — frictions, stats, validation (purged CV), regime (HMM), decay, capacity, cross-impact, promotion gates, sealed bundles
Engine	rift_engine	Backtest (vectorized + event-driven), walk-forward, Monte Carlo, parameter sweep, Bayesian smart-optimize, signal indicators, strategy base class, TCA, attribution
Trade	rift_trade	Capability-tiered auth (T0/T1/T2/T3), order proposal, signed execution, kill switches, position recon, websocket lifecycle
Research	rift_research	run_research_pipeline() chains data → backtest → WF/MC → advanced validations → sealed bundle
Portfolio	rift_portfolio	Multi-strategy supervisor: correlation guard, VaR, pair trades, daemon coordination
API	rift_api	HTTP REST API server exposing state files for institutional dashboards and PMS integrations
Strategies SDK	rift_strategies_sdk	Scaffold (rift new <name>) + the trend_follow reference strategy
CLI	@nexstone/rift-cli (TS)	40+ commands, oclif-based, spawns Python engine via subprocess
MCP	(in CLI)	59 MCP tools wrapping the same engine via rift serve — for AI agent integration

---

Three usage patterns

1. From the terminal:

rift research trend_follow BTC 4h            # full pipeline
rift backtest trend_follow --pair BTC --tf 4h # just the backtest
rift sweep trend_follow --pair BTC --tf 4h    # parameter sweep
rift new my_strategy                          # scaffold a new strategy

2. From Python:

from rift_research.research import run_research_pipeline

result = run_research_pipeline("trend_follow", pair="BTC", interval="4h")
print(result["promotion_verdict"])   # 5/5 gates, PASS

3. From an AI agent (Claude Desktop / Claude Code / etc.):

Configure your AI client to launch RIFT's MCP server:

{
  "mcpServers": {
    "rift": { "command": "rift", "args": ["serve"] }
  }
}

The AI now has access to 59 tools — backtest, walk-forward, sweep, scout, manual_trade, portfolio_start, audit_export, etc. Everything runs locally on your machine.

---

Vision principles

A few you should know before working in the codebase:

1. Frequency-agnostic — works on tick → 1m → 1h → 1d → 1w. No bar size baked into the math.
2. Style-agnostic — HFT MM, stat arb, momentum, carry, swing — all first-class.
3. Strategy-agnostic core — the engine does not know what strategy is running. Every promotion gate, every analytic phase, runs on any strategy.
4. HL-only forever — Hyperliquid is the single venue; no multi-exchange abstraction.
5. OSS ships everything except data and strategies — no tier gating, no feature flags.
6. Build it right the first time — no MVP / phase-N / "for now" framing in the code.
7. Top-quant rigor is the ceiling, not the floor — Renaissance-grade math runs underneath a Lightroom-grade interface.

The complete principle list lives in RELEASE.md.

---

Documentation

Start here:

• docs/QUICKSTART.md — 10 minutes from clone to your first sealed research bundle
• docs/INSTALL.md — full install instructions for macOS, Linux, WSL2

Using RIFT:

• docs/strategies/AUTHORING.md — write a strategy from scratch with the SDK
• docs/signals/AUTHORING.md — add your own custom scout signals
• docs/research/METHODOLOGY.md — what walk-forward / Monte Carlo / purged CV / DSR / alpha decay actually do
• docs/mcp/SETUP.md — wire RIFT into Claude Desktop / Claude Code / Cursor as an MCP server
• docs/CLI_REFERENCE.md — every command, every flag (auto-generated from --help)

Technical references:

• docs/AUTH_AND_EXECUTION.md — three-layer key model, capability tiers, audit substrate
• docs/BACKUP_AND_STATE.md — what's in ~/.rift/, what to back up, how to move installs
• docs/RUNBOOK_ALGO_MONITORING.md — daily / weekly checks for a long-running algo daemon
• engine/SIGNALS.md — indicator + signal reference (38 signals across 9 categories)

Project-level:

• CHANGELOG.md — what changed in each release
• KNOWN_ISSUES.md — current limitations and sharp edges
• PRIVACY.md — what data leaves your machine (spoiler: only Hyperliquid + AWS S3)
• SECURITY.md — how to report vulnerabilities
• CONTRIBUTING.md — dev setup, branch model, PR conventions
• CODE_OF_CONDUCT.md — community expectations
• RELEASE.md — release procedure and pre-release checklist

---

Contributing

This is an OSS project published by Nexstone. See CONTRIBUTING.md for the full guide. Quick reference:

# Run the full test suite
uv run --project engine pytest

# Run only fast tests (skips ones that need synced data)
uv run --project engine pytest -m "not slow and not mainnet"

# Build the TS CLI
cd packages/cli && pnpm build

Before opening a PR:

• All tests pass: pytest -m "not slow" returns 0
• New strategies have a promotion_gates config that honestly reflects what the strategy is (a slow trend-follower should declare min_trades=25, not lie about being institutional)
• New substrate primitives include unit tests in the same package

---

License

See LICENSE at repo root.