AFI-Reactor ⚡

AFI-Reactor is the canonical orchestrator for Agentic Financial Intelligence (AFI).

It orchestrates a multi-agent DAG-based pipeline capable of generating, analyzing, validating, executing, and observing financial signals at scale.

🤖 Droid Instructions

For AI agents and automated contributors: See AGENTS.md for canonical repo constraints, allowed tasks, and safe patch patterns.

Note: If AGENTS.md conflicts with this README, AGENTS.md wins. Critical: Read AFI_ORCHESTRATOR_DOCTRINE.md before modifying DAG logic.

🔒 Security

IMPORTANT: This repository uses environment variables for sensitive configuration.

Quick setup:

cp .env.example .env
# Edit .env with your credentials (NEVER commit this file)

---

🚀 Flexible DAG Architecture

AFI-Reactor implements a flexible, plugin-based DAG pipeline that can be customized with as many nodes as needed while adhering to AFI standards. The system supports dynamic pipeline construction through a composable plugin architecture.

Core Node Types

AFI-Reactor uses two categories of nodes: core nodes and plugin nodes.

Core Nodes (Required)

Core nodes are always present in the DAG and handle fundamental pipeline operations:

• AnalystNode — Loads analyst configuration, initializes enrichment pipeline, aggregates enrichment results (including AI/ML predictions), scores signals using ensemble ML models, and generates narratives
• ExecutionNode — Aggregates enrichment results, validates enrichment results, generates final scored signal, and prepares signal for observer
• ObserverNode — Observes the final scored signal, logs execution metrics, publishes signal to downstream consumers, and adds trace entries for execution tracking

Plugin Nodes (Optional & Composable)

Plugin nodes provide specific functionality and can be enabled/disabled as needed:

• ScoutNode — Scouts for new signals from external sources or AFI-native models, discovers trading opportunities, submits signals to enrichment pipeline, and tracks signal submissions for reward attribution
• NewsNode — Fetches news data from news providers, extracts news features, and stores news enrichment results
• SentimentNode — Fetches sentiment data from sentiment providers, calculates sentiment scores, and stores sentiment enrichment results
• PatternRecognitionNode — Detects chart patterns, calculates pattern metrics, and stores pattern recognition enrichment results
• SignalIngressNode — Ingests external signals, normalizes signal format, and stores signal ingress results
• TechnicalIndicatorsNode — Calculates technical indicators and stores technical indicator enrichment results
• AiMlNode — Calls AI/ML providers for predictions and stores AI/ML enrichment results (conviction scores, direction, regime, risk flags)

Pipeline Flow

1. Scout nodes execute first (independent signal sources, no dependencies)
2. Signal Ingress nodes execute second (may depend on Scout)
3. Enrichment nodes execute in parallel where possible (based on dependencies)
4. Required nodes execute last (analyst → execution → observer)

Note: This flexible architecture allows analysts to configure custom pipelines by selecting and ordering plugins as needed, as long as they follow the AFI Orchestrator Doctrine.

✅ Codex Health: 100% ✅ DAG Success Rate: 100% ✅ Agent Readiness: 100%

---

🧠 Node Architecture

Core vs Plugin Nodes

Core nodes:

• Always present in the DAG
• Handle fundamental pipeline operations
• Cannot be disabled or removed
• Execute in fixed order (analyst → execution → observer)

Plugin nodes:

• Optional and composable
• Can be enabled/disabled
• Can be ordered as needed
• Execute based on DAG configuration and dependencies

Signal Providers (Scouts)

Scouts are signal providers that bring strategies producing buy/sell signals or trade setups:

• ScoutNode — Discovers signals from external sources or AFI-native models
• Does NOT perform scoring (that's Analyst's responsibility)
• Does NOT enrich signals (that's Enrichers' responsibility)
• Tracks submissions for reward attribution (important for third-party Scouts)

Enrichment Layers

The pipeline supports multiple enrichment layers that can be configured:

• Technical: Technical indicators (RSI, MACD, EMA, etc.)
• Pattern: Chart pattern recognition (head and shoulders, triangles, etc.)
• Sentiment: Market sentiment analysis (social media, news sentiment)
• News: News analysis and feature extraction
• AI/ML: AI/ML predictions (conviction scores, direction, regime, risk flags)

---

⚡ Quick Start

Run the full validation pipeline:

npm run validate-all

---

📊 CI & Codex

Artifacts from CI include:

• config/dag.codex.json / config/ops.codex.json / config/schema.codex.json → canonical orchestrator config
• codex/codex.replay.log.json → generated replay log (runtime output; gitignored)
• tmp/dag-simulation.log.json → Simulation telemetry
• tmp/mentor-evaluation.json → MentorChain readiness scores

All commits to main trigger a full CI run with:

• DAG Replay Validation
• Signal Simulation
• MentorChain Evaluation
• Artifact Upload for 30 Days

---

📜 AFI Orchestrator Doctrine

afi-reactor is the ONLY orchestrator in AFI Protocol. All canonical pipelines, DAGs, and routing logic live here.

The 10 Commandments

1. afi-reactor is the orchestrator of AFI - All canonical pipelines, DAGs, and routing logic live here—not in afi-core, not in random helpers.

2. afi-core is our runtime library, not our boss - ElizaOS and AFI agents run inside pipelines defined by afi-reactor.

3. The DAG is law - Every signal path (ingest → enrich → score → mint/review) must be expressible as a Reactor DAG; ad-hoc flows are anti-patterns.

4. Agents are nodes, not gods - Individual agents (validators, mentors, tools) are pluggable nodes the DAG calls; they never control global orchestration.

5. Eliza's native orchestrator is an implementation detail - We may wrap or reuse it, but only as a node/operator under afi-reactor's authority.

6. State & replay belong here - Pipeline state, Codex replay, audits, and deterministic re-runs are owned by afi-reactor, even if storage is elsewhere.

7. Configuration is externalized - Reactor reads network, persona, and pipeline configs from afi-config and related registries; no hard-coded magic.

8. No token/econ logic in afi-reactor - Emissions, rewards, and AFI token rules live in afi-token; Reactor just emits events/hooks.

9. No infra glue in afi-reactor - Deployment, Terraform, K8s, etc. live in afi-infra / afi-ops. Reactor exposes clean interfaces they can target.

10. If orchestration logic doesn't fit this doctrine, it's in the wrong repo - Move it or refactor it until afi-reactor remains the single, boringly-obvious brain.

Full doctrine: See AFI_ORCHESTRATOR_DOCTRINE.md

---

🌌 AFI-Reactor Vision

AFI-Reactor is agent-first, modular, and framework-agnostic, powering use cases for:

• Retail traders and institutions
• Agent developers and ML researchers
• Real-time financial signal generation and execution
• Experimentation, stress testing, and open innovation

We embrace stress-tested resilience, inviting contributors to push the boundaries and make AFI stronger with every iteration.