projects.md

# Projects

## ZScape (Zandronum Server Browser) (Active, Jan–Feb 2026)

**Goal:** Build a modern, dark-themed Zandronum server browser to replace Doomseeker and Internet Doom Explorer.

**Naming:** Named "ZScape" after discussion Feb 1. Other contenders: Styx, Ztyx, ZScout.

**Implementation:** Vibe-coded with Opus in VSCode. Raptor Mini wrote documentation.
- Dark theme WinForms C# targeting .NET 10
- Concurrent hash verification (1 sec vs 10 sec for Doomseeker on 1GB+ mod list)
- Concurrent WAD downloads (unlimited downloads, 8 concurrent domains)
- Configurable concurrency limits
- Huffman encoding/decoding for Zandronum protocol
- Proper master server queries with segmented response handling
- Settings persistence to settings.json
- Verbose logging and hexdump debugging

**Status:** Working. Raymond can download files and join servers. Considering open-source but hesitant. Not actively discussed in recent chats.

**Key files:** README.md, SPECIFICATION.md uploaded Jan 29.

## Doom ML Map Generator (Active, Nov 2024–2026)

**Goal:** Train a machine learning model to generate playable Doom maps (specifically for Complex Doom/Zandronum).

**Recent Developments (Jan 20–27, 2026):**
- **Raycast pathfinding approach attempted but failed.** Raymond spent hours designing a 360-degree ray system with Info → Use → Explore → Traceback phases, but ultimately it didn't work.
- **Pivoted to A* pathfinding** for map validation. Opus implementing.
- **Still facing fundamental issue:** A* treats player as a 1x1x0 point, can fit through any gap and walk under closed doors without keys. Need to add collision detection (player hitbox is 32x32x56 units).
- Raymond refined phase order before pivot: RayCast (once per iteration) → Info → Use (auto-skip if nothing) → Explore (merged Traceback). Merged Traceback into Explore because both did the same thing (find best next location). RayCast now only casts rays once per iteration; Info/Use/Explore read from those results. Huge optimization.
- **Previous MAP02 bug:** Opus repeatedly thought the map was beatable without red key because abstract sector traversal didn't account for player hitbox or physical barriers. This is the same underlying issue the A* approach must solve.
- **Opus struggles:** Despite being the most intelligent consumer model, Opus repeatedly failed to grasp the problem until given visual evidence and explicit scrapping of its previous ideas. Finally making progress with the physics-based approach.
- **VSCode/Opus extension stable** – Raymond is not updating to avoid breaking it (Microsoft broke it multiple times in Jan). "2026 hasn't started well for AI services."
- **Unified progress framework** still desired – to discuss with Opus.
- Raymond feels he's finally making progress after 14 months.
- **Markov chain approach considered (Jan 26):** Simpler than deep learning, tracks state transitions and probabilities. Could work for coherent local structure in Doom maps. Not yet implemented.

**Tools:** TorchSharp (C#), CUDA on RTX 3060 Laptop (500 epochs in 42 seconds). VSCode with GitHub Copilot/Opus (cheap access to Opus for coding).

**Next steps:** Get A* pathfinding working with proper collision detection. Then integrate with reinforcement learning to validate maps during training. Possibly try Markov chain approach.

**Status:** Not actively worked on during ZScape focus, but planned to resume.

## Hedz Reverse Engineering
- 1998 game; reverse engineering 126,000+ lines of assembly to C#.
- Tools: IDA Pro, Ghidra, Cheat Engine, VS Code with GitHub Copilot.
- Built custom VSCode extension for method stack tracking.
- Currently ~400 methods left to verify against disassembly. Slow progress due to ZScape focus.

## Voxel Engine
- Started as a break from Hedz; C# (likely Unity or custom engine).
- Scale: 10 blocks per 1 player height.
- Inactive.

## Other
- ZScape is primary project focus as of Feb 2026.