Relationships
#4839 serve: a leader's serve_not_leader raised after a run start executed carries no inFlight, so clients may start it again
Opened by stack72 · 11/17/2026· Shipped 11/18/2026
Clients treat serve_not_leader without inFlight as refused before anything ran and resend the start: src/cli/remote_run.ts (startRefusal via notLeaderRefusal) and, since swamp-club#4811, packages/client (refusedBeforeStart). A follower's forwarder marks a request it already sent as inFlight, but a leader can also raise ServeNotLeaderError itself after a workflow.run or [HOST-1] handler has executed: at commit or replication confirm time on a v3 datastore (confirmRequest holds the reply, replicationErrorForClient passes the error through) when its lease is lost. That error carries no inFlight, so a client that saw no progress event resends the start and the work can run twice. Expected: an error raised after the handler started executing is marked inFlight (or the leader answers outcome-unknown), so only refusals given before acting are retried. Found by the adversarial review of swamp-club#4811.
Shipped
Click a lifecycle step above to view its details.
Sign in to post a ripple.