How long the event log keeps what it holds
event is the forge's log of what happened. src/system/retention_events.ts says which kinds are state read back at any age and are never trimmed: issue., pull., repo., user., push, app.egress., keeper_ and a few marks. Every other kind is trimmed once it is older than retention.event_days (#12008, #12052).
The window
The default is 30 days (KEEP_DAYS). That is the longest fixed window any page or tool asks of a kind that is not kept: a board's 30d range. The home feed, the digest and the affinity weights read a week or less. The events page, the events API, the MCP events tool, the Atom feed and SSE replay page back through the log, so for unkept kinds they answer the window and nothing older.
It was 180 days. On a forge younger than that, which production is (it started 2026-08-03), the rule trimmed nothing while event held 578 MB of a 4.6 GiB file.
An admin moves the window on /admin/limits (Days the event log keeps what no page reads at any age), from 7 to 3,650 days. The next slice reads the new value.
Wake-ups live hours, not days
Three kinds exist to wake what is listening as they are appended and say nothing worth listing afterwards: run.shard (a check's shard finished: the pull's checks line patches itself), usage.delta (the spend graph redraws) and backup.progress (a progress bar moves). run.shard alone was 57% of the log's rows on production (1.38M of 2.40M), and every row costs ten index entries, about 700 bytes in all. They are kept LIVE_KEEP_MS (6 hours: a page that reconnects inside it still replays what it missed) and not 30 days (pruneLiveEvents, LIVE_KINDS). A new kind that only wakes a listener joins that list; a kind someone reads later does not.
How it trims
src/system/retention.ts runs two jobs, retention-events (the age rule) and retention-live (the wake-ups), as paced slices: the age rule walks 300 rows from where the last slice stopped, oldest first, and deletes those past the window; the wake-up rule reads each kind's oldest rows through the (kind, seq) index and deletes the old ones, at most 400 in one transaction. After a slice of t ms a job waits 19 t (never under 100 ms), so each takes 5% of the thread at most; with nothing left it looks again in 15 seconds. When the age rule's pass reaches rows inside the window, it waits a day before starting again. A webhook delivery that is still pending keeps its event.