The main idea is to hook into all your coding agents (Claude Code, Codex, Cursor, etc...) and "swallow" their outputs after they finish tasks. This way, if you switch agents, the next one gets a richer, more complete prompt from Swallow ( something like that ).
Swallow acts as a planning agent. It "spits" the next plan and "swallows" completed tasks.
Another idea is to ingest everything (like Pac-Man) from your directories and files and add it to the project context.
You can check out the repo here: https://github.com/vule022/swallow
I tend to switch between coding agents a lot, I rarely stick to just one. I really dislike the "continue" or "resume conversation" flow in most agent CLIs, among other things
I’d love to hear your thoughts on this (Please ignore any naming choices ).