mirror of
https://github.com/vitorpamplona/amethyst.git
synced 2026-10-05 19:28:25 +00:00
Claude
53de512559
A migration moves every group's MLS state, cursor, draft, read position and KeyPackages to the new phone — and left the conversations behind. Every other piece has a second source: MLS state re-derives from the coordinator's stream, a KeyPackage can be republished. A cordn message has none. It is readable exactly once, at ingest, because both seal keys are epoch-derived and the cursor this very document carries has already advanced past everything behind it. A device seeded without the messages arrives holding every group and no conversation, permanently. Carried as `amethystMessages` on the group document, beside the other additive `amethyst*` fields, so an older reader ignores it rather than failing. Bounded per group by a byte budget, newest first. These blobs go to hosts whose limits we do not know, and a handoff that fails outright because one group is chatty is a worse outcome than one that carries a deep but bounded history. Budgeted in bytes rather than messages because a single long message can cost as much as a hundred short ones. Written before the cursor on import, for the same reason the live path writes them in that order: a seeding that saved the cursor and then failed would leave a device holding a cursor past a conversation it never wrote, with no way to ask for it again. My first version had the comment saying that and the code doing the opposite. Mutation-checked: a document that silently drops the messages fails the round-trip test and nothing else. Backup still excluded, per the plan — that call is still open. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012BfD4txdnsaPRXmNXbup9n