diff --git a/server/internal/services/patch_runs.go b/server/internal/services/patch_runs.go
index e997cf0..20f9679 100644
--- a/server/internal/services/patch_runs.go
+++ b/server/internal/services/patch_runs.go
@@ -145,6 +145,46 @@ func setServer(ctx context.Context, runID, serverID, from string, set bson.M) (b
 	return res.MatchedCount > 0, nil
 }
 
+// setServerForCommand is setServer scoped to one command: a late result from
+// an older, already-superseded command (say, a fresh dispatch to the same
+// server after a failed send) must not land on the new attempt.
+func setServerForCommand(ctx context.Context, runID, serverID, commandID, from string, set bson.M) (bool, error) {
+	fields := bson.M{}
+	for k, v := range set {
+		fields["servers.$."+k] = v
+	}
+	res, err := db.Col(patchRunsCol).UpdateOne(ctx,
+		bson.M{"run_id": runID, "servers": bson.M{"$elemMatch": bson.M{"server_id": serverID, "status": from, "command_id": commandID}}},
+		bson.M{"$set": fields})
+	if err != nil {
+		return false, err
+	}
+	return res.MatchedCount > 0, nil
+}
+
+// claimServerForDispatch atomically moves one server from queued/waiting_offline
+// to patching, guarded on the run still being running and not cancelled: a
+// tick that loaded the run before CancelPatchRun wrote cancelled_at must not
+// dispatch to it.
+func claimServerForDispatch(ctx context.Context, runID, serverID, from, commandID string, now time.Time) (bool, error) {
+	res, err := db.Col(patchRunsCol).UpdateOne(ctx,
+		bson.M{
+			"run_id":       runID,
+			"status":       models.PatchRunRunning,
+			"cancelled_at": bson.M{"$exists": false},
+			"servers":      bson.M{"$elemMatch": bson.M{"server_id": serverID, "status": from}},
+		},
+		bson.M{"$set": bson.M{
+			"servers.$.status":     models.PatchSrvPatching,
+			"servers.$.command_id": commandID,
+			"servers.$.started_at": now,
+		}})
+	if err != nil {
+		return false, err
+	}
+	return res.MatchedCount > 0, nil
+}
+
 func loadRun(ctx context.Context, filter bson.M) (*models.PatchRun, error) {
 	var run models.PatchRun
 	err := db.Col(patchRunsCol).FindOne(ctx, filter).Decode(&run)
@@ -191,18 +231,34 @@ func advanceRun(ctx context.Context, runID string) {
 	}
 	for _, tr := range patchrun.Advance(*run, now, connected) {
 		if tr.Dispatch {
+			// Each server's claim and any reset-after-failed-send get their
+			// own fresh, short-lived context rather than sharing the caller's:
+			// advanceRun can dispatch to many servers in one tick, each ack
+			// waiting up to Dispatcher's own timeout, and a caller-supplied
+			// context (StartPolicyRun's patchCtx, or a leader context that
+			// AdvancePatchRuns loses partway through) would otherwise expire
+			// mid-loop and silently drop a later server's writes.
 			cmdID := uuid.New().String()
-			ok, err := setServer(ctx, runID, tr.ServerID, tr.From,
-				bson.M{"status": models.PatchSrvPatching, "command_id": cmdID, "started_at": now})
-			if err != nil || !ok {
+			claimCtx, claimCancel := patchCtx()
+			ok, err := claimServerForDispatch(claimCtx, runID, tr.ServerID, tr.From, cmdID, now)
+			claimCancel()
+			if err != nil {
+				log.Printf("patch run %s: server %s: claim: %v", runID, tr.ServerID, err)
+				continue
+			}
+			if !ok {
 				continue
 			}
 			if err := dispatchPatch(tr.ServerID, cmdID, *run); err != nil {
 				// The agent dropped between the connection check and the send.
 				// Back to waiting: the next tick tries again while the window
 				// is open.
-				_, _ = setServer(ctx, runID, tr.ServerID, models.PatchSrvPatching,
-					bson.M{"status": models.PatchSrvWaitingOffline, "command_id": ""})
+				resetCtx, resetCancel := patchCtx()
+				if _, rerr := setServerForCommand(resetCtx, runID, tr.ServerID, cmdID, models.PatchSrvPatching,
+					bson.M{"status": models.PatchSrvWaitingOffline, "command_id": ""}); rerr != nil {
+					log.Printf("patch run %s: server %s: reset after failed dispatch: %v", runID, tr.ServerID, rerr)
+				}
+				resetCancel()
 			}
 			continue
 		}
@@ -305,7 +361,7 @@ func RecordPatchResult(instanceID, serverID string, r *pb.PatchResult) {
 		if updated.FinishedAt != nil {
 			set["finished_at"] = *updated.FinishedAt
 		}
-		if ok, _ := setServer(ctx, run.RunID, serverID, models.PatchSrvPatching, set); ok && updated.Status == models.PatchSrvRebooting {
+		if ok, _ := setServerForCommand(ctx, run.RunID, serverID, r.CommandId, models.PatchSrvPatching, set); ok && updated.Status == models.PatchSrvRebooting {
 			LogEvent(instanceID, "patch.reboot", "schedule", serverID, "",
 				fmt.Sprintf("%s rebooting for patch policy %s", s.Hostname, run.PolicyName))
 		}