Module narrated_events 

Remembers which events have already been narrated, so one background launch produces one ambient narration.

WHY THIS IS PROCESS-WIDE AND NOT A LOCAL IN register_handlers. That function runs once per WebSocket CONNECTION, so a local set is reset by every reconnect — and a reconnect is exactly when duplicates arrive. An honored truncate clears the frontend’s own nodeIdSet and re-parses the transcript, re-emitting every node as new, which re-invokes the narrate RPC for ids already narrated before the drop. Nothing else backstops it: the AmbientGateway singleton only guards calls that are concurrently in flight for a key, and releases its guard as soon as the first completes. Living on AppState alongside dock_snapshots and pending_background_pids is what makes the “narrate once” claim actually true. (reagent P1 on #3169.)

Bounded, unlike a bare HashSet: this accumulates one entry per narrated event for the lifetime of the process, which for a long-lived instance running many background tasks is unbounded growth for a guard that only needs recent history. Oldest entries are evicted first — a reconnect replays recent nodes, not ancient ones, so forgetting the distant past costs nothing.

Structs

Inner 🔒

NarratedEvents

Constants

MAX_REMEMBERED 🔒

How many narrated event ids to remember.