When running long AI coding tasks, prompt history eventually gets bloated, attention drops, and code quality falls off a cliff. I built LoopTroop to focus on context engineering rather than letting context rot happen.

Used a lot of codex do build it (even right now, lol).

Instead of letting a single model write code in one massive, noisy prompt, it takes a slow and precise approach:

• LLM Council planning: Multiple model passes draft, review, and anonymously vote on execution plans before touching any code.
• Context engineering: Each task step runs in a fresh context window with only the code it needs, avoiding bloated histories.
• Ralph loops: If a step fails tests or checks, a retry loop runs with clean context and targeted failure logs.
• Human in the loop: You can review plans, steer execution, and approve steps on a modern Kanban board GUI.

Just passed 100 stars on GitHub and I’m working on it daily.

Any feedback is more than welcome. If you give it a try and it works or doesn’t work for your setup, drop a note, happy to chat about it.