Collaborative editing

When multiple peers share a workspace, edits to the same state need a merge strategy. Hive's implemented model is event-level last-write-wins, with a character-level CRDT for concurrent free-text editing tracked as a follow-up.

Event-level convergence (implemented)

Every state mutation is a SessionEvent wrapped in a signed SessionEventEnvelope. The append-only envelope log is the source of truth, and the projector folds events deterministically — so replicas that have seen the same set of envelopes converge to the same state regardless of arrival order. Events are deduped by eventId, so restarts and double-sends are harmless.

For most fields this is exactly right: titles, role assignments, runtime selection, reactions, proposal create/vote, roster changes. Collisions on these are rare or low-stakes, and the deterministic event order resolves them.

The chat transcript is append-only at the envelope level — concurrent messages from different peers all land; there's no in-place edit to conflict over.

Concurrent free-text editing (roadmap)

True character-level concurrent editing of a single text field (e.g. two people typing into the same proposal body at once, Google-Docs style) needs a CRDT layer on top of the envelope log — a Replicated Growable Array for text plus a cursor/presence overlay. That is a tracked follow-up and not wired in the current build. Until it lands, proposal bodies use the event-level model above.