Thanks for this. I'd rather keep the coercion than throw: id is typed string, and String() normalizes the rare non-string at the boundary — the same single, predictable-type approach I use elsewhere. And remove() reads task.id, which is already that coerced string, so there's no asymmetry to guard against. Declining for that reason, but I appreciate the report.

(For what it's worth, v4 adds a TaskTimerError with a stable err.code for the errors the library does throw.) Closing.