For those using multiple bonded lines, proposed minimum reliability level guard mechanism in agent reporting to clueless. If packet loss is detected by a monitoring agent reporting to clueless, excluding the case where it is just link overload due to excess traffic, a filter mechanism monitoring loss plus traffic load shuts off that link by kicking it out of the bonded set. Then while in this state, the bad link is constantly LCP-pinged, and possibly also ICMP-pinged too, repeatedly to continue measuring packet loss in order to detect if the problem gets cured. This would of course need to be a state machine. Suggest that the criterion for link readmission has to be a fairly long period of zero packet loss.
[Also: Suggest also a second ‘outer’ state machine that monitors transitions into and out of the kicked-out/disabled state. This could make decisions aimed at optionally prevent looping - ie cycling in and out of the disabled state, if a problem is sporadic or a loss rate is marginal, right on or near a trigger/detection threshold.]
This feature would protect eg TCP from poor performance due to packet loss where a bad line causes far more problems than it is worth, seeing as packet loss is always very detrimental to overall transport performance.