When I was working on my own project, I felt that I was perhaps too pedantic about how stale-if-error was handled. I wanted something that clearly communicated whether a returned value was fresh or just fallback stale data because something upstream failed. I want people to think of this library for its cache policies, not its storage. What it leaves out, and where that goes, is the "Not included" list in the README.
Motivation has the backstory.
freshFor + max(staleWhileRevalidate, staleIfError). The windows are checked again on every read, so a store that keeps entries longer, or a store whose expiry is off, does not change what a read returns.MemoryStore evicts the least recently used entry past maxEntries; the read decision never looks at whether a store still has room.action-error event.Repletes doesn't care what the key looks like. A URL, a UUID, whatever: it all works the same. Only the namespace and version are prefixed onto it.
If you're using content-addressed keys (where the key is derived from the content), that's on you. The upside: the entry can never go stale, since the content determines the key. The downside: Repletes has no way to detect a collision if you hash wrong. That risk is yours now, not ours.
putConcurrent misses on the same key each call the action. Only the background refresh behind a stale read is deduplicated, one per key. Coalescing in-flight work belongs in the layer that also owns retries and deadlines, which is Firefly.
There is no way to prime the cache directly. Every stored value is the result of an action that ran, so an entry's windows always describe a call that completed.