Idea: integrate Ralph Loop best practices

[pedrosimao]

One thing that I found frustrating with the original Amazon Kiro IDE is that if I want to let the AI build multiple Tasks at once, Kiro tends to put all in one chat session, and keep on compacting context to open new sessions. We know that bigger context produces lower quality output.
So, I would normally need to start a new session for each sub-task... and surveil the context usage to be sure Kiro is not exploding it.

So, I would propose to create a ralph-loop-execute mode where Kiro will start new chat sessions for each sub-tasks and create new sessions till every goal is achieved.
It would be more of a hands off mode, where context usage is well managed.

[Koroqe]

Love this idea. Context bloat is the silent killer of output quality — I noticed the same thing where long sessions produce progressively worse code.

I built something similar: each sub-task runs as a fresh agent session with its own focused context, and a scratchpad file carries over just the essential state between sessions. Keeps each agent sharp without losing continuity.

https://github.com/Koroqe/claude-code-sdlc — the session persistence + sub-agent architecture might be relevant here.

[pedrosimao]

Love this idea. Context bloat is the silent killer of output quality — I noticed the same thing where long sessions produce progressively worse code.

I built something similar: each sub-task runs as a fresh agent session with its own focused context, and a scratchpad file carries over just the essential state between sessions. Keeps each agent sharp without losing continuity.

https://github.com/Koroqe/claude-code-sdlc — the session persistence + sub-agent architecture might be relevant here.

Nice stuff!
I opened a PR here:
#147

You can try it.
There are some scripts that measure context usage during the session, and the agent automatically stops the session if context is too large, always starting fresh.