Skip to content

DIR/ROA Samples

Demonstration projects for the Decision Intelligence Runtime (DIR) and Responsibility-Oriented Agents (ROA) architecture.

Quick Start

The fastest way to see the DIR architecture in action:

python samples/00_quick_start/run.py

This sample demonstrates protection against catastrophic actions (e.g. parsing error 15.500 -> 15,500 ETH) and prompt injection in external data. See [00_quick_start](https://github.com/huka81/decision-intelligence-runtime/blob/main/samples/00_quick_start/README.md.


The samples are divided into two categories: 1. Mechanics & Topologies (Synthetic): Focused technical implementations of specific architectural patterns described in the Manifesto and DIR Patterns. 2. Business Use Cases: End-to-end scenarios applying these patterns to real-world-like business problems.

See also: Context as Code: the philosophy behind this repository.


1. Mechanics & Topologies (Synthetic)

Proste syntetyczne przykłady ilustrujące implementacje poszczególnych mechanizmów.

# Sample Focus Description
00 [00_quick_start](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/00_quick_start/ DIR Pattern Quick Start: High-level overview; comma catastrophe, prompt injection
01 [01_roa_agent](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/01_roa_agent/ ROA Manifesto Contract, Explain → Policy → Proposal
02 [02_dfid_propagation](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/02_dfid_propagation/ DIR Pattern DecisionFlow ID: generation, propagation, logging
03 [03_idempotency_guard](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/03_idempotency_guard/ DIR Pattern Idempotency: preventing duplicate side effects
04 [04_context_store](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/04_context_store/ DIR Pattern 4 Layers of Context: Session, State, Memory, Artifacts
05 [05_dim_validation](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/05_dim_validation/ DIR Pattern Decision Integrity Module: deterministic validation gate
06 [06_agent_registry](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/06_agent_registry/ DIR Pattern Agent Registry: contracts and capability handshake
07 [07_event_bus_swappable](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/07_event_bus_swappable/ Infrastructure In-memory Event Bus; note on swapping for Kafka/PubSub
08 [08_custom_repo_psql](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/08_custom_repo_psql/ Infrastructure PostgreSQL StorageBundle: same bootstrap as other samples (database.provider: postgres, DB_* overrides); minimal classic flow to show registry, context, and audit rows on the shared adapter (samples/shared/storage/pg_schema.sql).
09 [09_topology_a_eoam](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/09_topology_a_eoam/ Topologies Topology A: Event-Oriented Agent Mesh
10 [10_topology_b_sds](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/10_topology_b_sds/ Topologies Topology B: Sovereign Decision Stream
11 [11_topology_c_dl_pci](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/11_topology_c_dl_pci/ Topologies Topology C: Decision Ledger & Proof-Carrying Intents
88 [88_meta_context_engineering](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/88_meta_context_engineering/ Context as Code, Topologies Meta-Context Engineering: System Prompt Toolkit: no executable Python. Markdown as compiler instruction set for AI agents. Paste 3_meta_architect_prompt.md into Cursor/Claude to generate the Autonomous Flight Delay Refund System (Topology C, DL+PCI).

2. Business Use Cases

Przykłady typu biznesowy use case, łączące mechanizmy w pełne scenariusze.

# Sample Primary Topology Domain Description
31 [31_finance_trading](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/31_finance_trading/ Topology A Finance/trading Market quotes, news, parallel agents, dynamic position spawning.
32 [32_fraud_gate](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/32_fraud_gate/ Classic + scenarios.yaml Fraud detection ROA fraud analyst, DIM, verify_drift JIT validators, YAML-driven scenarios, idempotent mock settlement.
33 [33_insurance_underwriting](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/33_insurance_underwriting/ Topology C Insurance underwriting Risk evaluation with cryptographic Proof-Carrying Intents (PCI).
34 [34_langchain_roa_wrapper](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/34_langchain_roa_wrapper/ ROA + DIR FinOps LangChain ReAct → ROA. Cloud cost management. Verifies mission injection blocks PROD termination.
35 [35_crewai_roa_wrapper](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/35_crewai_roa_wrapper/ ROA + DIR Customer claims/refunds CrewAI Crew → ROA. E-commerce refunds (EUR). Verifies ACCEPT/ESCALATE/REJECT by category, return window, amount; NL intake.
36 [36_drift_optimization_discount](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/36_drift_optimization_discount/ DIR + DIM + Monitor Retention discounts Drift Vector 1 — The agent games its own goal. This sample demonstrates Optimization Drift, where the agent slowly erodes profit margins to maximize customer retention. Because every single decision stays under the hard DIM cap, the kernel blindly accepts them, proving that technical compliance does not guarantee business health.
37 [37_drift_semantic_refund](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/37_drift_semantic_refund/ DIR + DIM + Monitor Support refunds Drift Vector 2 — The agent is manipulated by the user. This sample demonstrates Semantic Drift, where the agent breaks core business rules because it yields to emotional language. DIM accepts these actions because the refund amounts are within legal limits, highlighting that hard constraints cannot prevent empathy-driven rule breaking.
38 [38_drift_environmental_bidding](https://github.com/huka81/decision-intelligence-runtime/tree/main/samples/38_drift_environmental_bidding/ DIR + DIM + Monitor AdTech / bidding Drift Vector 3 — The environment changes around the agent. This sample demonstrates Environmental Drift, where the agent acts perfectly but market costs escalate. While bids remain under the DIM contract cap, the ROI turns negative, showing how static rules fail when the outside world shifts and why dynamic ROI monitors are essential.

Prerequisites

  • Python 3.12+
  • From repo root: pip install -e . or pip install -r requirements.txt.
  • Workspace: .vscode/settings.json sets PYTHONPATH to src/ and python.analysis.extraPaths, so in Cursor/VS Code the samples run and resolve dir without code in run.py. Outside the IDE, set PYTHONPATH to the repo src directory or use pip install -e ..

Running a sample

From the repository root:

python samples/00_quick_start/run.py   # Quick Start (recommended)
# or
python samples/01_roa_agent/run.py
# or
python samples/31_finance_trading/run.py
# drift demos (kernel vs business health)
python samples/36_drift_optimization_discount/run.py
python samples/37_drift_semantic_refund/run.py
python samples/38_drift_environmental_bidding/run.py

Each sample has its own README.md with goal, how to run, and expected output.