Most AI coding assistants are fundamentally broken. Tools like Claude Code and OpenCode eagerly load 10k+ tokens of monolithic system prompts into a single context window before you even type. This guarantees context amnesia and destroys local inference speeds.
I built late to fix this. It’s a single-binary Go/BubbleTea TUI built around a subagent orchestrator pattern. The main prompt is strictly ~100 lines. The orchestrator routes tasks to transient subagents with isolated context windows, explicitly preventing the "Debugging Decay" (https://arxiv.org/abs/2506.18403) and context pollution seen in monolithic agents.
Architectural notes:
* Linux Native: Strictly respects XDG base directories. No Windows support. No padded Electron GUI.
* Execution Load: The rapid context switching breaks the mainline llama.cpp scheduler. You must compile the specific PR linked in the README (or just point it at any standard API via localhost:8080).
* Licensing: BSL 1.1 to keep the core orchestrator logic out of VC-backed wrappers.
I also open-sourced pure-go-sgd (AGPLv3) today for those interested in bare-metal Go infrastructure: https://github.com/mlhher/pure-go-sgd
