ReDSL

Refactor + DSL + Self-Learning — AI-Native DevOps & Refactoring OS

⚠️ This is not a typical requirements DSL. This is an autonomous operating system for AI-driven software engineering.

ReDSL is an experimental framework combining LLM, formal runtime DSL, CI/CD, and self-refactoring loops into one autonomous code lifecycle system.

Current project state

Based on the 2026-04-09 code2llm analysis:

• Files: 114
• Functions: 781
• Classes: 112
• Lines of code: 19,151
• Average complexity: CC̄ = 4.1
• Critical hotspots: 3
• Duplications / cycles: 0 / 0
• Test suite: 468 collected tests
• Next refactor: split format_cycle_report_markdown(), format_batch_report_markdown(), and LLMLayer.call()

🧠 What ReDSL Really Is

Not just a requirements DSL. This is an AI-driven software lifecycle system:

🔥 KEY PARADIGM SHIFT

❌ Before (typical DSL): describe requirements → generate documentation → manual interpretation

✅ Here: describe system → system has CI/CD, tests, linting, deployment, refactor pipeline → LLM can intervene in code

This is an autonomous development system.

🏗️ Architecture: Autonomous Loop

SUMD → DOQL → taskfile → pyqual → testQL → LLM refactor loop → deployment

Detailed Flow:

┌─────────────────────────────────────────────────────────────────────┐
│                    AUTONOMOUS CODE LIFECYCLE                         │
│                                                                       │
│   🧾 SUMD ──► ⚙️ DOQL ──► 🔄 taskfile ──► 🧪 pyqual ──► 🤖 LLM       │
│       │          │           │            │           │               │
│       ▼          ▼           ▼            ▼           ▼               │
│   [Spec]    [Runtime]   [DevOps]    [Quality]   [Refactor]            │
│       │          │           │            │           │               │
│       └──────────┴──────────┴────────────┴───────────┘                │
│                          │                                            │
│                          ▼                                            │
│   ┌─────────────────────────────────────────┐                          │
│   │  REFACTOR ORCHESTRATOR                 │                          │
│   │  PERCEIVE → DECIDE → PLAN → EXECUTE    │                          │
│   │       ↓                    ↓          │                          │
│   │  REFLECT → REMEMBER → IMPROVE        │                          │
│   └─────────────────────────────────────────┘                          │
│                          │                                            │
│                          ▼                                            │
│   ┌─────────────────────────────────────────┐                          │
│   │  VALIDATION LAYER                      │                          │
│   │  regix (regression) │ vallm │ sandbox   │                          │
│   └─────────────────────────────────────────┘                          │
│                          │                                            │
│                          ▼                                            │
│   [deployment] ◄── CI/CD ◄── quality gates ◄── auto-PR              │
│                                                                       │
└─────────────────────────────────────────────────────────────────────┘

🚨 WHAT'S REALLY NEW HERE

🧠 A. "Code as controllable system"

This is NOT: code + AI This is: operating system for code

🔁 B. Self-learning loop

You have: tests, lint, quality gates, refactor pipeline, LLM model 👉 A system that can self-correct its own code

🧩 C. DSL = control interface

DOQL is not a declarative language - it's a system orchestrator:

workflow[name="test"] {
  run pytest
}

Features

• 🔍 Static Analysis - Integration with popular linters and metrics tools
• 🧠 LLM with Reflection - Generate refactoring proposals with self-reflection loop
• ⚡ Hybrid Engine - Direct refactorings for simple changes, LLM for complex ones
• 📊 DSL Engine - Define refactoring rules in readable YAML format
• 💾 Memory System - Learn from refactoring history (episodic, semantic, procedural)
• 🚀 Scalability - Process multiple projects simultaneously
• 🔄 Autonomy Loop - Perceive → Decide → Plan → Execute → Reflect → Memory Update

Installation

pip install redsl

Refactor a single project (dry run)

redsl refactor ./my-project --max-actions 5 --dry-run

Refactor without dry run (apply changes)

redsl refactor ./my-project --max-actions 10

Get output in YAML format (for integration)

redsl refactor ./my-project --format yaml

Get output in JSON format (for APIs)

redsl refactor ./my-project --format json

# Process semcod projects with LLM
redsl batch semcod /path/to/semcod --max-actions 10

# Hybrid refactoring (no LLM) for semcod projects
redsl batch hybrid /path/to/semcod --max-changes 30

# Batch processing with JSON output
redsl batch semcod /path/to/semcod --format json

Every refactor and batch run also writes a Markdown report next to the project or root folder:

• redsl_refactor_plan.md — --dry-run output
• redsl_refactor_report.md — executed refactor cycle
• redsl_batch_semcod_report.md — batch summary for batch semcod
• redsl_batch_hybrid_report.md — batch summary for batch hybrid

Analyze code quality

redsl pyqual analyze ./my-project

Analyze with custom config

redsl pyqual analyze ./my-project --config pyqual.yaml

Get analysis in JSON format

redsl pyqual analyze ./my-project --format json

Apply automatic fixes

redsl pyqual fix ./my-project

# Check configuration
redsl debug config --show-env

# View DSL decisions for a project
redsl debug decisions ./my-project --limit 20

GitHub Actions example

• name: Run reDSL analysis run: | redsl refactor ./ --max-actions 5 --dry-run --format yaml > refactor-plan.yaml

• name: Upload refactoring plan uses: actions/upload-artifact@v3 with: name: refactor-plan path: refactor-plan.yaml

# Use with jq for JSON processing
redsl refactor ./ --format json | jq '.refactoring_plan.decisions[] | select(.score > 1.0)'

# Pipe to file for review
redsl refactor ./ --format yaml > review-plan.yaml

# Extract only high-impact decisions
redsl refactor ./ --format yaml | yq '.refactoring_plan.decisions[] | select(.score > 1.5)'

Environment Configuration

Create .env file:

# LLM Configuration
OPENAI_API_KEY (set in your environment)
REFACTOR_LLM_MODEL=openai/gpt-4
REFACTOR_DRY_RUN=false

# Custom settings
REFACTOR_MAX_ACTIONS=20
REFACTOR_REFLECTION_ROUNDS=2

Simple Actions (no LLM)

• REMOVE_UNUSED_IMPORTS - Remove unused imports
• FIX_MODULE_EXECUTION_BLOCK - Fix module execution blocks
• EXTRACT_CONSTANTS - Extract magic numbers to constants
• ADD_RETURN_TYPES - Add return type annotations

Implementation note: the deterministic AST helpers now live in redsl/refactors/ast_transformers.py, and redsl.refactors plus redsl.refactors.direct re-export them for backward compatibility.

Complex Actions (with LLM)

• EXTRACT_FUNCTIONS - Extract high-complexity functions
• SPLIT_MODULE - Split large modules
• REDUCE_COMPLEXITY - Reduce cyclomatic complexity

Fresh-project smoke test

To quickly verify that ReDSL runs in a brand-new project, create a tiny temporary project and run the CLI in dry-run mode:

mkdir -p /tmp/redsl-smoke
cat > /tmp/redsl-smoke/main.py <<'PY'
import os


def main() -> None:
    return None


main()
PY

python3 -m redsl analyze /tmp/redsl-smoke
python3 -m redsl refactor /tmp/redsl-smoke --dry-run --max-actions 5

REST API

Start the API server:

# Using uvicorn directly
uvicorn redsl.api:app --reload --host 0.0.0.0 --port 8000

# Using the CLI
redsl api --host 0.0.0.0 --port 8000

Refactor a Project

curl -X POST "http://localhost:8000/refactor" \
  -H "Content-Type: application/json" \
  -d '{
    "project_path": "./my-project",
    "max_actions": 5,
    "dry_run": true,
    "format": "json"
  }'

Batch semcod processing

curl -X POST "http://localhost:8000/batch/semcod"
-H "Content-Type: application/json"
-d '{ "semcod_root": "/path/to/semcod", "max_actions": 10, "format": "yaml" }'

Hybrid batch processing

curl -X POST "http://localhost:8000/batch/hybrid"
-H "Content-Type: application/json"
-d '{ "semcod_root": "/path/to/semcod", "max_changes": 30 }'

# Get configuration
curl "http://localhost:8000/debug/config?show_env=true"

# Get decisions for a project
curl "http://localhost:8000/debug/decisions?project_path=./my-project&limit=10"

Analyze code quality

curl -X POST "http://localhost:8000/pyqual/analyze"
-H "Content-Type: application/json"
-d '{ "project_path": "./my-project", "format": "json" }'

Apply fixes

curl -X POST "http://localhost:8000/pyqual/fix"
-H "Content-Type: application/json"
-d '{ "project_path": "./my-project" }'

### Interactive API Documentation

Once the server is running, visit:
- Swagger UI: http://localhost:8000/docs
- ReDoc: http://localhost:8000/redoc

## ⚖️ Markdown + AI vs ReDSL — Comparison

### 📝 Markdown + AI
- Input: loose text
- AI interprets
- No hard system structure
- **This is an assistant**

### 🧠 ReDSL (this project)
- Input: structural system (SUMD + DOQL)
- AI operates in a **controlled runtime**
- Has quality pipeline + CI/CD + refactor loop
- **This is an autonomous system for managing code lifecycle**

| Criterion | Markdown + AI | ReDSL |
|-----------|---------------|-------|
| **UX** | ✅ Wins | ⚠️ Complex |
| **Adoption** | ✅ Easy start | ⚠️ High entry cost |
| **Simplicity** | ✅ Intuitive | ⚠️ Many abstractions |
| **System control** | ❌ None | ✅ Deterministic runtime |
| **Lifecycle automation** | ❌ Manual | ✅ Auto-pipeline |
| **CI/CD + AI integration** | ❌ None | ✅ Native |
| **Determinism** | ❌ Non-deterministic | ✅ DSL-driven |

**Conclusion**: Markdown + AI wins in productivity and UX. ReDSL wins **only if** AI development becomes fully autonomous and companies accept "DSL system as devops runtime".

## 📊 Project Assessment

| Criterion | Score | Justification |
|-----------|-------|---------------|
| 🧠 **Innovation** | **9/10** | Close to Devin, Auto-refactoring systems, AI CI/CD pipelines |
| ⚙️ **Technical coherence** | **8.5/10** | Full dev pipeline, quality system, docker + CI + LLM |
| 🚧 **Practical adoption** | **6/10** | Very complex, high entry cost, no market standard |
| 📉 **Risk** | **High** | Many DSL abstractions, LLM dependency, no production usage proof |

### 🎯 FINAL CONCLUSION

👉 **This is NOT a "requirements DSL" anymore**

👉 **This is: an experimental operating system for AI-driven software engineering**

- ❌ Not a typical DSL
- ❌ Not competition to Markdown (different category)
- 🟢 This is **AI-native DevOps + refactoring OS**
- 🟡 Very ambitious, but hard to deploy

## Architecture

┌─────────────────────────────────────────────────────┐ │ ORCHESTRATOR │ │ (loop: analyze → decide → refactor → reflect) │ ├─────────────┬──────────────┬────────────────────────┤ │ ANALYZER │ DSL ENGINE │ REFACTOR ENGINE │ │ ─ toon.yaml│ ─ rules │ ─ patch generation │ │ ─ linters │ ─ scoring │ ─ validation │ │ ─ metrics │ ─ planning │ ─ application │ ├─────────────┴──────────────┴────────────────────────┤ │ HYBRID REFACTOR ENGINES │ │ ─ DirectRefactorEngine (no LLM) │ │ ─ LLM RefactorEngine (with reflection) │ ├─────────────────────────────────────────────────────┤ │ LLM LAYER (LiteLLM) │ │ ─ code generation ─ reflection ─ self-critique │ ├─────────────────────────────────────────────────────┤ │ MEMORY SYSTEM │ │ ─ episodic (refactoring history) │ │ ─ semantic (patterns, rules) │ │ ─ procedural (strategies, plans) │ └─────────────────────────────────────────────────────┘

## Configuration

Environment variables:
- `OPENAI_API_KEY` or `OPENROUTER_API_KEY` — API key
- `REFACTOR_LLM_MODEL` — LLM model (e.g., `openrouter/moonshotai/kimi-k2.5`)
- `REFACTOR_DRY_RUN` — test mode (`true`/`false`)

## Examples

| Directory | Description |
|-----------|-------------|
| `examples/01-basic-analysis/` | Project analysis from toon.yaml files |
| `examples/02-custom-rules/` | Define custom DSL rules |
| `examples/03-full-pipeline/` | Full cycle: analyze → decide → refactor → reflect |
| `examples/04-memory-learning/` | Memory system: episodic, semantic, procedural |