Vertical traversal via the navigation mesh, not the behaviour tree. Rather than write a "climb wall" task, I built a custom off-mesh navigation link actor that designers place in a level. It registers its entry and exit with the navmesh as a connection the pathfinder can cost like any other edge. The result of this is that a plain "move to this location" call routes a bot up a wall automatically whenever doing so shortens the path. No AI wiring, no special cases, and level designers extend bot mobility by easily placing actors.

Facing and Detection. Bots were flickering between combat and chase states 40 transitions in a single measured session. The cause was that the AI controller overwrites control rotation every frame from its focus point, and the path-following system publishes a focal point along the path on every update, so anything a service wrote that frame was deleted. Worse yet, releasing facing whenever a branch deactivated meant a momentary line-of-sight blip swung the bot away, which dropped the target out of its sight cone, which dropped sight again. I moved facing onto the engine's focus system at a priority that outranks path following, and reduced the release of facing to one place: the single point at which a target stops existing. A grace period bridges the natural sight flicker around cover, which the previous debounce would never cover.

Failure handling that does not punish the team. A climb that times out blacklists that link for that bot only, for 15 seconds, so it re-routes instead of walking back to the same wall. One bot's bad luck must not deny a good route to everyone. It takes two failures inside a 30-second window, because a single fluke should not cost a working route. Repeated blacklist entries at one link are a level-authoring signal, which is exactly what the my logging suite is designed to surface.

Three-strike stuck recovery. A bot is stuck only when it both failed to move and failed to close on its goal. The second test exists because a bot strafing around a target barely changes position but is not stuck. So it escalates: re-path, then break contact with a jump and a lateral step, then abandon the goal entirely. The third strike is there for a reason: a bot carrying an objective is hard-blocked from climbing, so if it gets routed over cover, abandoning the goal is

its only escape.

The Jump Script. The navmesh bakes jump links over cover, but with no smart-link object attached, so the engine creates the connection and then never propels anything across it leading to bots appearing stuck on waist high cover. I wrote the crossing behaviour myself, splitting at chest height a mantle, driving climb input into the obstacle so the character's existing mantle behaviour (which I also authored) can find the edge; above that it is a jump; below 20 cm it is a drop-down that is just walked off. Four things had to happen for either to work, and the second pass fixed all of them (suspend crowd steering, set an aim override, feed movement input every frame, and add a one-shot forward boost on leaving the ground, because bots were landing back on the near side).

Crowd avoidance. I swapped in the crowd-following component so bots avoid each other, then suspended crowd steering for the duration of every traversal. The crowd manager only skips *falling* agents, and a wall climb is a custom movement mode, so otherwise two systems write velocity simultaneously. The suspend resets on possession so a bot that dies mid-climb cannot respawn permanently un-steered.