#!/usr/bin/env python3
"""
Single-file pipeline: Extract, normalize, and load trending RSS data into SQLite
"""
import sys
import sqlite3
import urllib.request
import xml.etree.ElementTree as ET

def fetch_and_index_feed(feed_url: str, db_path: str = "feed.db"):
    req = urllib.request.Request(feed_url, headers={'User-Agent': 'Mozilla/5.0'})
    with urllib.request.urlopen(req) as response:
        xml_data = response.read()

    root = ET.fromstring(xml_data)
    conn = sqlite3.connect(db_path)
    cursor = conn.cursor()

    cursor.execute("""
        CREATE TABLE IF NOT EXISTS articles (
            id TEXT PRIMARY KEY,
            title TEXT,
            link TEXT,
            published_at TEXT
        )
    """)

    for item in root.findall('.//item'):
        title = item.find('title').text if item.find('title') is not None else 'Untitled'
        link = item.find('link').text if item.find('link') is not None else ''
        pub_date = item.find('pubDate').text if item.find('pubDate') is not None else ''

        cursor.execute("""
            INSERT OR REPLACE INTO articles (id, title, link, published_at)
            VALUES (?, ?, ?, ?)
        """, (link, title, link, pub_date))

    conn.commit()
    print(f"Indexed feed into {db_path} successfully.")

if __name__ == "__main__":
    if len(sys.argv) > 1:
        fetch_and_index_feed(sys.argv[1])
    else:
        print("Usage: python index_feed.py <RSS_URL>")

flowchart TD
    A["Untrusted Web Input (HTML / Email / PDF)"] --> B["Quarantined Reader LLM (Zero Tool Access)"]
    B --> C["Structured JSON Summary (Strictly Validated Schema)"]
    C --> D["Privileged Orchestrator LLM (Has Database / Write Access)"]
    D --> E["Safe Execution Output"]

1/4 🧠 Why @simonw is the sharpest mind documenting the AI engineering shift:
He observed something profound: after decades of living inside Python REPLs and Jupyter notebooks, senior developers have almost completely replaced them with AI coding loops.

2/4 The economics of software changed forever:
The marginal cost of writing one-off data extraction and exploration scripts has dropped to zero. If you need a custom parser, you prompt an agent, inspect the 40 lines of code, run it, and delete it.

3/4 But beware of Indirect Prompt Injection:
The moment you give your AI agent access to browse the live web or read emails, malicious text can hijack its instructions. Simon's dual-LLM sandboxing model is essential reading for AI developers.

4/4 Follow @simonw for daily, highly practical deep-dives on LLMs, SQLite, prompt defense, and tool design.