2026-06-01 Meeting Minutes

Completed Work

• Ran the Godot project to check the map generation results.
• Identified an issue where the map screen appeared differently than expected.
• Customized web-search-mcp from a personal private repository to fit the desired workflow.
• Tested Qwen Code, Codex, and OpenCode.
• Connected and tested the gemma-4-e4b-it.F16.gguf family models on RunPod.
• Reviewed Hugging Face’s Godot‑specific model bekoozkan/godot-gemma-4-e4b-it-GGUF because several models tended to suggest past Godot code patterns.

Observed Issues

• Qwen Code had a major problem of repeating the same answers and plans.
• Qwen Code did not improve the repetition issue easily even when the model was changed unless the setup matched the documentation.
• Codex and OpenCode lacked understanding of the Godot context and image interpretation.
• General models tended to suggest code patterns from older Godot versions rather than the latest.
• In map improvement tasks that required image verification, the model alone could not make reliable judgments.
• Some tools failed while planning and used sub‑agents or research flows inappropriately.
• Apart from Qwen Code, the usability of the web-search MCP was not as good as expected.

Decisions

• Separate today’s notes into a retrospective and meeting minutes.
• Write the retrospective in docs/retrospectives/2026-06-01.md.
• Write the meeting minutes in docs/meetings/2026-06-01.md.
• Add a document classification and a summary for June 1, 2026 to the README.
• Apply changes directly to main without creating a new branch.
• Starting tomorrow, gather Godot work patterns and failure cases to attempt fine‑tuning ourselves.

Notes

The actual project work today centered on running Godot and checking the map state. The rest of the time was mostly spent validating AI coding tools, MCP, model connections, repetitive answer issues, and image interpretation problems.