Built a terminal coding agent in Go — would rather have it torn apart than upvoted

https://preview.redd.it/decuuxdjlbrh1.png?width=3444&format=png&auto=webp&s=3df87269c27e354727aae9416c9e52f25dc59f37 Tapioca v1.2.0, MIT, solo project. Same space as Claude Code, OpenCode and Crush those are what it's inspired by. I wrote it because I wanted one I could actually shape: panels, keybinds, themes, per-tool permission rules and hooks around tool calls are all config, not whatever the vendor picked. It's also largely built agentically, most of the code was written by LLM agents with me directing and reviewing. Worth stating up front, because it's the reason I want outside eyes: this thing has had a lot of machine review and exactly one human's. The model reads, writes and runs things in your repo, every mutating call behind a permission prompt. Runs against Ollama or llama.cpp with no key and nothing leaving localhost; hosted providers work too. So the permission model, the sandboxing, the provider layer, the Go itself. Pull on whatever looks weakest. https://github.com/juacamole/tapioca