Have you ever actually seen the agent's context? What if you could fully control it?

ljedrz — Fri, 02 Oct 2026 11:01:58 +0000

Most terminal LLM agents become black boxes when context fills up. You can't inspect what the model is carrying, remove specific items, or know why something was dropped.

kamchatka is different: a Rust terminal agent where context, tools, permissions, and requests are explicit, first-class state.

The context tab shows every item, its cost, whether it's in the next request, and why not. You can move items between active/elided/excluded, pin things compaction can't touch, and see the exact request.

Most agents optimize for autonomy. kamchatka optimizes for legibility.

Repo: https://github.com/ljedrz/nachalnik

I lied to an LLM in its own memory. It caught the lie and fixed it.

ljedrz — Tue, 01 Sep 2026 14:38:33 +0000

https://ljedrz.github.io/nachalnik/

DEV Community: ljedrz

Have you ever actually seen the agent's context? What if you could fully control it?

I lied to an LLM in its own memory. It caught the lie and fixed it.