AlienCyborgModernTempleOS 🛸🤖

An automated multi-agent orchestration system to look for alien signals in YouTube recommendation algorithms (pattern analysis).

This is a local, offline-friendly ASP.NET Core Razor Pages app that:

1. Accepts a ZIP of saved YouTube Home HTML snapshots (ex: SingleFile.html saves).
2. Extracts video “tiles” (title + thumbnail URL when available).
3. Runs a multi-agent LLM pipeline (Signal → Interpreter → Skeptic → Synth).
4. Writes outputs to disk and displays them in a clean Activity panel.

---

Features

• ✅ Upload ZIP → safe extraction (blocks zip path traversal)
• ✅ Parse tiles from:
  • ytInitialData (when present), and/or
  • DOM parsing fallback (SingleFile-friendly)
• ✅ Multi-agent orchestration:
  • Signal: clusters + anomalies (JSON output)
  • Interpreter: roleplay hypothesis grounded in evidence
  • Skeptic: normal explanations + bias checks
  • Synth: final combined report with confidence + next experiments
• ✅ Activity endpoint that auto-refreshes and shows:
  • Final report
  • Signal JSON
  • Interpreter output
  • Skeptic output
• ✅ Results persisted per job under App_Data/jobs/<jobId>/

---

Project layout (high-level)

• Pages/Index.cshtml – Upload UI + Activity panel (auto-refresh)
• Pages/Index.cshtml.cs – Upload handler, ZIP extraction, parsing, orchestration, file writes
• AlienOrchestrator.cs – Multi-agent pipeline + proof-text capping
• LmStudioChatClient.cs – OpenAI-compatible chat client for LM Studio
• LlmAgent.cs, IAgent.cs – Agent wrapper(s)
• AgentPrompts.cs – Prompt “system” roles for each agent
• YouTubeSnapshotParser.cs – HTML parsing → VideoTile extraction

---

Requirements

1) .NET

• .NET SDK 8 or newer (recommended: latest LTS)

2) LM Studio (local LLM server)

You need LM Studio running a local OpenAI-compatible server:

• Install LM Studio
• Download/load a chat model
• Start the Local Server (OpenAI-compatible)

This project expects the server at:

• http://localhost:1234/v1/chat/completions

If you change the port or base URL, update your app configuration accordingly.

3) (Optional) AngleSharp

If you are using DOM fallback parsing, install:

• AngleSharp NuGet package

---

Configure the app

You can configure LM Studio settings using appsettings.json (recommended) or hard-code in LmStudioChatClient.

Example appsettings.json snippet:

{
  "LmStudio": {
    "BaseUrl": "http://localhost:1234/v1/",
    "Model": "your-model-name-here"
  }
}

Then bind in Program.cs (or inject via options). If you already inject LmStudioChatClient and AlienOrchestrator, you’re good.

---

Run it

1) Start LM Studio server

• Load your model
• Start the local server
• Confirm it answers at http://localhost:1234/v1/chat/completions

2) Run the web app

dotnet restore
dotnet run

Open the printed URL (usually https://localhost:xxxx).

---

How to use

1. Create YouTube Home snapshots:

   • Open YouTube Home (signed in if you want personalization)
   • Save each page as a SingleFile HTML (or similar)
   • Do this multiple times (ex: yt_home_01.html … yt_home_20.html)

2. ZIP the HTML files together.

3. Upload the ZIP in the app.

4. Watch the Activity panel:

   • It will show state changes (extracted → running_agents → completed)
   • Expand accordions for Signal / Interpreter / Skeptic / Final Report

---

Where results go

Each upload creates a job folder:

App_Data/jobs/<jobId>/
  snapshot.zip
  extracted/
    (your html files)
  status.json
  results.json

• status.json contains state + counts
• results.json contains the serialized JobResult (agent outputs)

---

Troubleshooting

“400 Bad Request” + context overflow

If your ZIP is large, your input can exceed the model’s context length.

Fix options:

• Reduce the number of tiles/pages you feed into the proof text
• Cap the proof text length (recommended)
• Use a model with a larger context window
• Increase context settings in LM Studio (if supported by your model/runtime)

“found 0 tiles” / “Could not locate ytInitialData”

This can happen if:

• Your snapshot doesn’t include the ytInitialData payload
• The page is saved in a way that strips scripts/data

Fix options:

• Enable the DOM fallback parsing path (AngleSharp)
• Ensure you saved a fully-loaded YouTube Home page
• Try SingleFile with “include scripts” enabled (if available)

Output is in Chinese / includes <think>

Some models ignore prompt instructions. Fix options:

• Switch models
• Add stronger “English-only / no reasoning tags” constraints
• Add a post-filter step to strip <think>…</think> blocks

---

Credits

• Built with ASP.NET Core Razor Pages
• Uses an OpenAI-compatible local server (LM Studio)
• Optional DOM parsing via AngleSharp