The workaround everyone uses is to re-explain the problem from memory, badly, at the start of every fresh session. It costs twenty minutes and it is still lossy — you remember the goal, you do not remember that the third approach failed for a subtle reason, so you watch the new agent walk into it again.

None of the existing answers fit. A static instructions file describes the repo, not what is happening in it. The agent's own context dies with the process and is compacted away long before that. Resume flags only work within one tool, one account and one machine — which is precisely the boundary you are trying to cross.

So the context has to outlive the session that produced it, be captured without anyone remembering to, and be served back small enough to actually read. That is the whole product.

The opinions that follow from it are stated plainly in the codebase: preserve what cannot be reconstructed and drop what can; a brief you cannot reproduce is a brief you cannot debug; a false conflict is worse than a missed one. They are the useful part.