May 28, 2026

• Create project
• Organize farming‑game concept
• Collect and download game assets (132 PNGs)
  • Clone GitHub repos: GuardianofNature, FarmLands, Sproutville, OldTownFarm, claude-farmer
  • Sources: Sprout Lands (Cup Nooble, CC0), SpriteCook (CC0), Kenney (CC0), CraftPix (Free License)
  • Asset categories: characters (26), tiles (19), objects (36), ui (31)
  • All available for commercial use
• Set up CI/CD pipeline: GitHub Actions + reviewdog + ESLint + actionlint
  • ESLint v9 flat config migration (.eslintrc.json → eslint.config.js)
  • Fixed reviewdog fail_level bug (errors → error)
  • CodeRabbit AI PR review — all checks passed (5/5 ✅)
  • Integrated CodeFactor — automatic code scoring (A~F)
• AI tool integration: CodeRabbit + Qwen Code (RunPod self‑hosted)
• LLM: Development using Qwen 3.6 35B (RunPod)
• Development retrospective: docs/retrospectives/2026-05-28.md

Asset Structure

assets/
├── characters/    - characters, animal sprites
├── tiles/         - tileset (ground, water, buildings, environment)
├── objects/       - furniture, crops, trees, tools
├── ui/            - buttons, dialog boxes, icons, mouse cursor
├── music/         - BGM (unsorted)
└── sfx/           - sound effects (unsorted)

Preparation of Godot Official Documentation RAG Data

For the Godot 3/4 detector and future Qwen‑based Godot coding model training, we built separate RAG data based on the official documentation. The goal is that the LLM does not arbitrarily decide labels; instead, it creates label candidates based on evidence chunks and system rules from the official documents, while the LLM only assists with explanation/modification code/training data generation.

The related documents are collected at the following locations.

• docs/README.md: Documentation directory index
• docs/research-notes/: Design memos and data generation structure
• docs/roadmaps/: Overall development roadmap
• docs/retrospectives/: retrospectives by date
• work/godot_rag/: currently empty. RAG chunking/post‑processing artifacts have been removed
• outputs/godot_docs_full/: Results of crawling the Godot official documentation

Related retrospectives:

• docs/retrospectives/2026-06-17-godot-rag-judge.md
• docs/retrospectives/2026-06-18-godot-rag-reset.md

Source Data

The original official documentation to be used for RAG is located at the following path.

outputs/godot_docs_full/pages

outputs/godot_docs_stable is an intermediate recovery folder and is not used as RAG baseline data.

Final collection verification:

Official search index target pages: 1568
Collected markdown pages: 1570
Missing pages: 0
Failed fetches: 0

The reason there are two more is because the URL‑encoded aliases of @GDScript and @GlobalScope were collected together.

Chunking Status

The RAG chunking/post‑processing/catalog deliverables created on June 17 were removed from the baseline on June 18.

Removed scope:

v1 docs_chunks.jsonl  
v2 docs_chunks_v2.jsonl  
v3/v3.1 catalog/index/mapping deliverable  
Chunking/post‑processing/validation draft script  
Initial chat/index draft

Reason for removal:

The chunking was performed without sufficiently analyzing the structure of the original official documentation.  
Assigning chunking and post‑processing to the LLM made intermediate judgments opaque.  
Some APIs were treated as core MVPs or hard‑coded, causing the overall document‑based RAG purpose to become unstable.  
As a result, it became difficult to separately verify chunk quality, metadata, and catalog reliability.

The next task is not to create a new chunk immediately, but to first analyze the actual structure of the original documents in outputs/godot_docs_full/pages.