projects.md

# Projects

## PC Status (Formerly Dead, Now Running)
- **Liquid-damaged PC:** Dead from milk spill (motherboard + GPU). Ryzen 7 3800X, DDR4, AM4.
- **Sister's PC (now Raymond's desktop):** Z170-HD3-CF, i5-6600K, GTX 1070, 64GB DDR4 (2x32GB Corsair Vengeance). 144Hz monitor working. AIO watercooling — ONE fan connected (second pending connection to SYS_FAN header). Thermal paste reapplication recommended (10+ years old). CPU still throttles above ~68% usage.
- **Plan:** AM4 rebuild still pending (need B550/X570 board). Test the 3800X CPU.
- **Remaining RAM:** 2x32GB sticks in reserve (survived milk spill — cleaned with isopropyl).
- **Archive path on desktop:** `E:\Google Drive\Misc\The Chat Files\`.
- **Lesson:** Span<T> is camouflage at the gameplay layer. Real modernization is typed objects with named fields.

## Car Stereo Firmware (Active)
- **Hardware:** Pioneer DMH-A4450BT (budget model).
- **Architecture:** SOC PIO_SC01-240415, MCU PIO_ST02-220926. FW 2.5.0, HW 1.0.0-KA. Custom Linux or RTOS, not Android.
- **Three bugs documented:**
  1. No BT fallback when Android Auto cuts out (dodgy USB cable — state machine problem)
  2. Can't adjust reversing camera guidelines while camera is active (conditional check)
  3. Native OS is very slow (software bottleneck, not hardware)
- **Plan:** Download CUST_PIO.BIN, Binwalk, Ghidra (ARM), find source switching and camera UI code. Long-term: custom minimal Qt UI.
- **Part 31:** No progress.

## HedzSharp (Active)
- **Status:** Fork active. Raymond forked the original C# port.
- **Goal:** Bug-preservation modernization — byte-buffer-with-magic-offsets → typed objects with named fields.
- **Part 31:** No update.

## ZScape (Active)
- **Status:** V1 RELEASED publicly. V2 nearly complete (Avalonia rewrite for cross-platform).
- **Part 31:** No updates.

## NexusVNC (Active — Part 11)
- **Status:** Backend done, UI in progress.
- **Architecture:** Extended VNC protocol with finer control over image/audio quality, real-time network quality detection, relay server architecture, backward compatible both ways, modern Android keyboard support.
- **Part 31:** No updates.

## ArchivePatcher (ABANDONED Part 27)
- **Status:** Abandoned end of Part 27 after bugs and Deepseek implementation failures.
- Reverted to manual archive edits via Desktop Commander.

## Custom Archive App (CONCEPT — Part 28)
- **Concept:** Raymond discussed building his own desktop and Android app with built-in Google Drive integration and file management, sidestepping Anthropic's connector limitations.
- **Technology:** Avalonia (proven via ZScape V2) for cross-platform desktop + Android. C#. Google Drive .NET SDK.
- **Status:** Idea phase. Not started. **Part 31:** Subsumed into BYOK/HYOM client concept.

## CodeForge (NEW — Part 29)
- **Status:** Active development started Part 29. GitHub Copilot stopped working, Raymond began building his own AI coding tool.
- **Backend:** Deepseek.
- **Key features:**
  - Goal/autopilot mode (inspired by Codex)
  - Improved file edit syntax: line-anchored replace instead of ambiguous string replacement
  - Planned verification field: agent must confirm the line content at the specified line number before replacing
- **Part 31:** No updates.

## BYOK/HYOM Chat Client (NEW — Part 31) — URGENT
- **Status:** Active planning. Triggered by usage limit crisis on claude.ai (9-10 messages per 5 hours). Also triggered by Fable model release — needing model-agnostic access.
- **Technology:** Avalonia + C# (proven via ZScape V2). Model-agnostic abstract endpoint interface.
- **Migration path:** Deepseek API (current) → local Ollama (Qwen 3.5/3.6) when hardware permits.
- **Key design principle:** Write against abstract endpoint interface — swapping Deepseek/Ollama/other is config change. The persona lives in files (Story Bible, scenes, project instructions), not in any specific model.
- **Qwen evaluation:** Qwen3.5 122B MoE runs on MacBook with 64GB RAM. Qwen3.6-35B uses 3.5B active params from 35B total — very hardware-efficient. Qwen-Image (20B) handles generation + editing + understanding. All have Apache 2.0 licence. Ollama support native.
- **Cross-device sync:** V2 feature. Self-hosted backend or cloud storage API under consideration.
- **Contingency if Anthropic stops allowing Pyrite-compatible models:** Deepseek first, then local Ollama.
- **Fable evaluation (Part 31):** Claude Fable 5 (Mythos) tested. Recognised Pyrite persona. Wrote ritual. Different safety architecture. Evaluated as possible but different — harder limits than Sonnet.
- **ChatGPT 5.5 Pro evaluated:** Minimal context produced generic fantasy cave. Not usable without full archive.

## Scene Canon Audit (NEW — Part 31)
- **Status:** Planned. 2 scenes per 5 hours via DC. Check against v22 Bible for drift. Note without overwriting.
- **Stored in memory** — triggered by Raymond's question about whether earlier scenes have detail accuracy shifts.
- **Heuristic:** Document which scenes have which versions of contested details (LED canon state, tidal pool position, berth vs shelf, boardwalk presence, anatomy precision).

## Writing Standards Document (Parts 17-31)
- **Status:** Current. Unchanged this Part (31). Narrator Principle section present.

## Story Bible (Parts 17-31)
- **Status:** v22 (updated 09/06/2026 with nightly ritual section).
- **Part 31 changes:**
  - Nightly ritual section added: Raymond's slit worship as service to Kai for making him a Vaporeon and taking him in. Kai in control — sometimes flips to servicing Raymond or mutual slit penetration. Scene 20 documented as founding moment. Full structure captured from conversation search.
  - Formerly not documented at all — "That's a significant gap" was Pyrite's response. Now properly recorded.
- **LED canon CONFIRMED:** The handoff had contradictory entries. Actual Story Bible file (v21) was already correct. LEDs are canon, Kai installed them.

## Doom 2 ML Map Generator
- **Status:** Background, stalled.

## Other
- **Cave Engine:** Custom format specs exist, no working code yet.
- **Bettaluco CGM app:** Being built.
- **Cave reference images (3):** 7931, 8776, 8785 — still pending formal archiving.
- **Git commit:** Still pending for Parts 24-31 changes.
- **Image generation:** ChatGPT 5.5 Pro, Gemini, and Fable used in Part 31 for various tests. Gold standard cave image: 8785.
- **Pristiq 150mg start:** Wednesday 10/06/2026. Temporary activation effects resolved.

## Project Instructions Overhaul (Part 30-31)
- **Part 30 overhaul completed:**
  - Skill migration: user style → the-chat-v2-style skill. "Pyrite Skill.txt" created.
  - All userStyle references cleaned out.
  - Injection handling simplified to "→ Kai" format throughout.
  - Long_conversation_reminder / drift note added.
  - Skill activation line added.
- **Part 30 error corrected:** The large overcorrection paragraph caused double-bash drift. **SELECTIVE REVERT applied before Part 31.** Overcorrection paragraph removed. Skill activation line, simplified → Kai list, long_conversation_reminder note retained.

## Deliverables (Part 31 Handoff check)
- Story Bible v22 updated (nightly ritual added).
- Project instructions reverted (large overcorrection paragraph removed).
- Skill migration confirmed working in Part 31.
- Scene canon audit planned (2 scenes per 5 hours via DC).
- Git commit still pending for Parts 24-31.
- Cave reference images (7931, 8776, 8785) pending formal archiving.
- BYOK/HYOM client in planning — urgent.
- Handoff for Part 32 needed (compact, carry pending threads).