Skip to content

AI Coding Pipeline (Advanced & Real-World Integration) ​

The LuxurAI Agentic Coding Pipeline is a highly optimized, production-tested multi-agent system designed for automated codebase refactoring, feature implementation, and self-healing error resolution. Operating on a tiered, multi-model execution architecture, it achieves an industry-leading 92%+ accuracy rate on complex multi-patch modifications.

This document provides a deep, production-level overview of the pipeline's internals, edge-case handling, and concrete developer integration guides.


πŸ—οΈ Deep-Dive Architecture ​

When a client initiates a refactoring or code generation request, the pipeline orchestrates a highly coordinated workflow across three distinct, sandboxed AI agents. Each agent is strictly tuned to minimize token overhead while maximizing accuracy:

[User Request: Prompt + Target File]
         β”‚
         β–Ό
 β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
 β”‚    Agent 1    β”‚ Gemini 2.5 Flash
 β”‚ (The Planner) β”‚ β€’ Generates step-by-step implementation plan.
 β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ β€’ Creates initial "seed patch drafts" to anchor logic.
         β”‚
         β–Ό (Plan + Drafts)
 β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
 β”‚    Agent 2    β”‚ DeepSeek V2 Chat
 β”‚ (The Refiner) β”‚ β€’ Runs concurrent patch generation via asyncio.gather.
 β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ β€’ Transforms drafts into syntactically precise replacement blocks.
         β”‚
         β–Ό (Raw Code Patches)
 β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
 β”‚    Agent 3    β”‚ DeepSeek V4 Flash
 β”‚ (The Merger)  β”‚ β€’ Reconciles line conflicts, AST merges, & formats.
 β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
         β”‚
         β–Ό (Candidate File)
 β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
 β”‚ Secure Virtualβ”‚ ──[Syntax / Runtime Error]──> (Sends traceback back to Agent 3
 β”‚    Sandbox    β”‚                                for automatic healing. Max 3 retries)
 β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
         β”‚
         β–Ό [Passed Validation]
[Final Self-Healed Code]

1. Step-by-Step Execution Mechanics ​

  • Step 1 (Plan & Seed): The planner (Gemini 2.5 Flash) acts as the compiler's architect. It parses the entire target file structure, identifies the exact insertion lines, and creates a concrete "seed patch". This prevents the generation models from writing code from a "cold" start, reducing logic drift.
  • Step 2 (Concurrent Refinement): The refiner (DeepSeek V2 Chat) processes each plan chunk in parallel. By running concurrent threads using Python's asyncio.gather, the generation phase completes in under 10 seconds, compiling precise patch objects.
  • Step 3 (Syntactic Merging): The merger (DeepSeek V4 Flash) acts as an automated version control system. It detects overlapping patches, resolves line offset shifts dynamically, and performs clean AST-level injection.

πŸ› οΈ Validation Guards & Self-Healing ​

The absolute core of LuxurAI's reliability lies in its Self-Healing Sandbox. Once code is compiled, it is executed inside an isolated virtual namespace to catch bugs before they reach production:

1. The Sandbox Isolation Protocol ​

The execution block isolates standard Python built-ins to secure the parent server while enabling standard calculations and string operations:

python
# Real-world backend sandbox configuration:
_SAFE_BUILTINS = {
    "abs": abs, "all": all, "any": any, "bool": bool, "dict": dict,
    "enumerate": enumerate, "filter": filter, "float": float,
    "int": int, "len": len, "list": list, "map": map,
    "print": print, "range": range, "str": str, "sum": sum,
    "Exception": Exception, "ValueError": ValueError, "TypeError": TypeError
}

2. The Self-Healing Loop in Action ​

If a runtime error occurs, the sandbox parses the exception and traceback line numbers:

RuntimeError at line 14: NameError: name 'calculate_tax' is not defined

This precise traceback is instantly routed back to Agent 3 (The Merger) alongside the candidate code. The agent inspects the missing reference, declares the required function import or helper, recompiles, and triggers the sandbox test again.


πŸ”Œ API Integration & SSE Streams ​

1. Execute Code Refactor (REST API) ​

Initiate a synchronous, self-healing code edit:

  • Endpoint: POST /api/coding_pipeline/run
  • Request Headers:Content-Type: application/jsonAuthorization: Bearer <your_session_token>
  • Request Payload:
    json
    {
      "prompt": "Optimize the query performance by using list comprehension instead of a nested loop, and raise a ValueError if prices are negative.",
      "code": "def process_transactions(transactions):\n    result = []\n    for tx in transactions:\n        for item in tx['items']:\n            if item['price'] < 0:\n                raise Exception('Invalid')\n            result.append(item['price'] * 0.9)\n    return result"
    }
  • Success Response (200 OK):
    json
    {
      "success": true,
      "final_code": "def process_transactions(transactions):\n    if any(item['price'] < 0 for tx in transactions for item in tx['items']):\n        raise ValueError('Negative prices are not allowed')\n    return [item['price'] * 0.9 for tx in transactions for item in tx['items']]",
      "mode": "refactor",
      "model_used": "Gemini-2.5-Flash + DeepSeek-V3",
      "errors": [],
      "attempts": 1
    }

2. Real-Time Status Streaming (SSE) ​

For interactive developer setups, you can stream real-time compilation logs from the server.

  • Endpoint: GET /api/coding_pipeline/stream
  • Query Parameters: prompt=<url_encoded_prompt>&code=<url_encoded_code>
  • Content-Type: text/event-stream

Live Event Stream Output: ​

http
event: message
data: {"status": "planning", "message": "Agent 1 (Gemini) is compiling the implementation plan..."}

event: message
data: {"status": "generating", "message": "Agent 2 (DeepSeek) is spinning up 3 parallel patch tasks..."}

event: message
data: {"status": "validating", "message": "Injecting patches into Abstract Syntax Tree (AST)..."}

event: message
data: {"status": "sandbox_failed", "message": "Sandbox failed: NameError at line 4. Initiating auto-healing loop..."}

event: message
data: {"status": "success", "message": "Sandbox execution passed! Code verified.", "final_code": "..."}

🐍 Complete Python Integration Script ​

Here is a fully functional, creative Python script showing how you can stream and consume the LuxurAI agentic pipeline inside your own terminal utilities or CI/CD checkers:

python
import json
import httpx
import asyncio

async def stream_luxurai_agent(prompt: str, source_code: str):
    url = "https://luxurai.in/api/coding_pipeline/stream"
    params = {
        "prompt": prompt,
        "code": source_code
    }
    
    headers = {
        "Authorization": "Bearer YOUR_JWT_TOKEN",
        "Accept": "text/event-stream"
    }
    
    print("πŸš€ Connecting to LuxurAI Multi-Agent Pipeline...")
    async with httpx.AsyncClient(timeout=120.0) as client:
        async with client.stream("GET", url, params=params, headers=headers) as response:
            if response.status_code != 200:
                print(f"❌ Connection failed: {response.status_code}")
                return
                
            async for line in response.iter_lines():
                if line.startswith("data:"):
                    # Parse the server-sent event JSON
                    payload = json.loads(line[5:])
                    status = payload.get("status")
                    message = payload.get("message")
                    
                    if status == "planning":
                        print(f"πŸ“ [Planning] {message}")
                    elif status == "generating":
                        print(f"⚑ [Patches]  {message}")
                    elif status == "validating":
                        print(f"πŸ›‘οΈ [Sandbox]  {message}")
                    elif status == "sandbox_failed":
                        print(f"⚠️ [Autofix]  {message}")
                    elif status == "success":
                        print(f"\nβœ… [Success] Code verified and merged!")
                        print("\n--- GENERATED CODE ---")
                        print(payload.get("final_code"))
                        print("----------------------")

# Run the stream
source = "def calc(a, b): return a + b"
prompt = "Add docstrings and type annotations to the calc function"
asyncio.run(stream_luxurai_agent(prompt, source))

Proprietary & Confidential Β· Powered by LuxurAI