iil-reflex 0.2.1

REFLEX — Reflexive Evidence-based Loop for UI Development

iil-reflex

REFLEX — Reflexive Evidence-based Loop for UI Development

Evidence-based UI development methodology with LLM-powered domain agent, UC quality checker, and failure classifier. Pure Python — no Django dependency.

Architecture

reflex/
├── agent.py       # DomainAgent (variable domain, LLM-powered)
├── quality.py     # UC Quality Checker (11 criteria, rule-based)
├── classify.py    # Failure Classifier (decision tree + LLM)
├── config.py      # ReflexConfig from reflex.yaml
├── providers.py   # KnowledgeProvider, DocumentProvider (Protocol)
├── types.py       # Dataclasses (Results, Questions, Entries)
└── templates/     # promptfw .jinja2 templates (6 templates)

Installation

pip install iil-reflex
# Optional: Playwright for Zirkel 2
pip install iil-reflex[playwright]

Quick Start

from reflex.agent import DomainAgent
from reflex.config import ReflexConfig
from reflex.quality import UCQualityChecker

# 1. Load hub config
config = ReflexConfig.from_yaml("reflex.yaml")

# 2. UC Quality Check (rule-based, no LLM needed)
checker = UCQualityChecker(config)
result = checker.check(uc_text="...", uc_slug="sds-upload")
print(f"Score: {result.score_percent}%, Passed: {result.passed}")

# 3. Domain Agent (LLM-powered)
agent = DomainAgent(
    config=config,
    llm=your_llm_provider,
    knowledge=your_knowledge_provider,  # optional
    documents=your_document_provider,   # optional
)
research = agent.research("SDS Upload Pipeline")
questions = agent.generate_interview(research)
kb = agent.distill_kb(research, expert_answers={...})

Hub Configuration (reflex.yaml)

hub_name: risk-hub
vertical: chemical_safety
domain_keywords: ["SDS", "CAS", "GHS", "REACH"]
quality:
  min_acceptance_criteria: 2
  max_uc_steps: 7
  require_error_cases: true
viewports:
  - {name: mobile, width: 375, height: 812}
  - {name: desktop, width: 1280, height: 800}
htmx_patterns:
  banned: ["hx-boost"]
  required_on_forms: ["hx-indicator"]
permissions_matrix:
  /substances/: {anonymous: 302, viewer: 200, admin: 200}

Provider Pattern

All external dependencies use Protocols (Dependency Inversion):

from reflex.providers import KnowledgeProvider, LLMProvider

class OutlineProvider(KnowledgeProvider):
    def search(self, query, limit=5):
        return mcp3_search_knowledge(query, limit=limit)

class GroqProvider(LLMProvider):
    def complete(self, messages, action_code=""):
        return groq_client.chat(messages=messages)

Mock providers included for testing: MockKnowledgeProvider, MockDocumentProvider, MockLLMProvider

REFLEX Methodology

Three quality circles — no artifact without evidence:

1. Zirkel 0 — Domain KB (DomainAgent + Expert sign-off)
2. Zirkel 1 — UC Quality (11 criteria, 100% score required)
3. Zirkel 2 — Playwright Tests (1 test per acceptance criterion)

Failure Classification:

• UC_PROBLEM → UC needs revision (Zirkel 1 restart)
• UI_PROBLEM → Wireframe needs fix
• INFRA_PROBLEM → Server/browser/network issue

Dependencies

• iil-promptfw>=0.7.0 — prompt template rendering
• pyyaml>=6.0 — config file parsing
• Optional: playwright — for Zirkel 2 test execution

Related

• ADR-162 — Full ADR
• iil-promptfw — Prompt framework
• Platform: achimdehnert/platform

License

MIT