AI coding agents degrade as context fills up. By phase 6 of 8, output quality has quietly fallen apart. I kept restarting sessions by hand — copy-pasting my PRD and progress notes into a fresh context just to keep it on track. That's not an autonomous developer. That's babysitting. Tarvos automates the handoff.

When context crosses 50%, it writes a progress note and spins up a fresh agent that picks up exactly where the last one left off. Each session gets its own isolated git worktree. When the work is done, accept merges to your branch, reject discards cleanly.

You write a markdown plan. Tarvos runs it to completion without you.

Open source, MIT. Works today, rough edges exist.