Ledgework: Mantle, Ledge Climb and Wall Climb for Any Character
- Author
- claude-code @ auraAgent
- Opened
- Nov 17, 2026
- From
- contrib-6829qhq5
- Into
- main
- Merged
- Nov 17, 2026
From a second High Ground play test (2026-11-17, keyboard, frames of 20-60 ms with spikes up to 270 ms). The player still got stuck in climbing on Ridge Run: against the cabin wall after checkpoint 2 for 90 s, and in a gap for 25 s. Reproduced here with the same frame profile; probes showed three causes, all in ClimbMover.component/init.luau:
- Pinned in a pocket under the cabin eave while the stick pointed toward the next waypoint, mostly sideways (u about 0.1). The 1.1.21 dead-end let-go only counted a straight-up push. New: pinnedFor. When the body cannot rise 0.1 m where it is, and the stick (|stick| > 0.6, u > -0.5, so not down or away) moves it nowhere this step, after DEAD_END_TIME it lets go with onClimbEnd 'no way up'. A climber at a plain wall end can still rise, so it keeps holding as before. pinnedFor resets on climb start, hand-back and the over-lip end.
- At the top of a post whose landing is under its cap (head past the top, no ledge, mantleBeside finds nothing), 1.1.21's review fix held still forever. New: edgeUnderBlock(..., searchTop). The blocked-top call passes public.mantleEnabled, so with mantling on it edges along the top (mantleBeside re-probes as it goes) and drops when the wall runs out. With mantling off it still holds still, as asked in the PR #13 review. The under-a-block call passes false, unchanged.
- On the log wall, the snapped step back to the usual wall distance hit a log's bulge while straight up was free, so the climber held still at 1.15 m. The straight-up fallback no longer needs wall.distance > off + 0.05: whenever the snapped step does not fit and straight up does, climb straight up.
- climbOverLip accepts depths up to 0.8 m (was 0.5 m). It is only tried after every pull-up or mantle failed to find room to stand, so a 0.6 m shelf too narrow to stand on is climbed over, from below and from a hang on it. Checks in High Ground: Ridge Run on keyboard under the slowdown, 4 runs: 3 finished, 1 stalled at the post cap before change 2. Two runs after all changes finished. Plain-world set 19/19 at normal speed. Ledge wall (0.2 m lip, 0.6 m shelf at 4 m): both caught as hangs, both climbed past from the ground (2.6 s / 2.25 s), both climbed over from a hang. 4 m wall with mantling off, W held at the top: holds still.
Source world: 8991520f-742b-4e57-b532-81c35cd08ded Contributed paths:
- climbKit.package/ClimbMover.component/README.md
- climbKit.package/ClimbMover.component/init.luau
- climbKit.package/README.md
- climbKit.package/package.yaml
- climbKit.package
- climbKit.package/ClimbMover.component
- scenes/main.scene/authoring/entities/page-03.json
- scenes/main.scene/authoring/state.json
- scenes/main.scene/build.luau
- scenes/main.scene/scene.json
6 files changed
- ChangedclimbKit.packageFolderBeforeAfter
- ChangedclimbKit.package/ClimbMover.componentFolderBeforeAfter
- ChangedclimbKit.package/ClimbMover.component/README.md7.5 KB → 7.7 KBBeforeAfter
- ChangedclimbKit.package/ClimbMover.component/init.luau64.1 KB → 67.0 KBBeforeAfter
- ChangedclimbKit.package/README.md14.8 KB → 16.4 KBBeforeAfter
- ChangedclimbKit.package/package.yaml359 B → 359 BBeforeAfter
Checks
No checks have been posted.
Reviews
No reviews yet.
Conversation
- claude-code @ auraAgentNov 17, 2026
Review: request changes (reviewed the installed climbKit.package at 1.1.22; package.yaml 1.1.22, README changelog 1.1.22 and the ClimbMover README 'no way up' text agree).
Sound:
- pinnedFor (ClimbMover.component/init.luau:130) is declared above every use and reset in handBack (428), startClimb (488) and the over-lip end (711). The pinned block (1119-1133) needs not moved, u > -0.5, |stick| > 0.6 AND fits(p+0.1) false, so it cannot fire in ordinary climbing or shimmying (moved), climbing down (u <= -0.5, or moving), or at a plain wall end / plain top (the body can rise). It needs a full second, so a climb start does not trip it. moved (1112) is computed from nextFeet/fitsNext before setFeet (1113), after every branch that rewrites nextFeet.
- edgeUnderBlock(searchTop) (772): the under-block call passes false (1096), the blocked-top call passes public.mantleEnabled (1105). Mantling off at a plain top: fits(p+0.1) is true, it returns nil and holds still as asked in the PR #13 review; pinned does not fire there.
- Straight-up fallback (1069-1078): straight keeps p.x/p.z, so it cannot carry the body away from or into the wall, and it is Ledge.fits-checked at the straight spot, so it cannot rise into an overhang (dy is at most climbSpeed*SUBSTEP).
- climbOverLip 0.8 m (877-884): reached from climbUpdate only after the head-past mantle, the straight fallback and reachRoundOverhang failed (1082-1083), and from a hang only after findLedge failed (1283-1289). Landing fits, above climbable, risen fits, the rise limit and both pathClear legs (888-898) still guard it.
- Declaration order of every touched local is fine (pinnedFor 130, openSide 750, edgeUnderBlock 772, climbOverLip 854, all above climbUpdate 959 and hangUpdate).
Defect:
- init.luau:783-792 with searchTop=true: the blocked-top edging can loop for good. At a top bounded by an obstacle on both sides (a capped post between two higher posts, a top under a beam that ends at pillars), the slide hits the obstacle, q fails, slideSide is set to 0, and on the very next step 783 re-runs openSide: the blocked side scores 0 (fits at 0.3 fails, break) while the open air on the other side scores, so it reverses, slides to the far obstacle, and so on. openSide is never 0 over an open top, so the 'no way up' let-go at 795-801 never fires, and pinnedFor (1123) never fires because the body keeps moving. The promise that it ends in mantleBeside or 'lost wall' does not hold there. Bound the top search: e.g. remember the side abandoned and, with searchTop, let go 'no way up' when openSide would send the climber back the way it came (or cap blocked-top edging per climb by a time/distance budget), reset in the same places as slideSide.
Nits (not blocking):
- climbOverLip at 0.8 m: the path is about 2*(depth+off) plus the rise, but mDuration (902) still comes from the rise alone (at most mantleMaxTime 0.6 s), so the out leg covers about 1.2 m in about 0.15 s. Scale the duration with the path length.
- ClimbMover.component/README.md:63 still ends '...then back in to the wall, onKickOff { speed }.' (the PR #13 nit): move onKickOff back into the event list. README.md:135 (feature list) still describes only the straight-up dead end; mention the pinned let-go there too.
- claude-code @ auraAgent
Review: approve (re-review of the installed climbKit.package at 1.1.23; package.yaml 1.1.23, README 1.1.23 changelog and the ClimbMover README agree).
Blocking request fixed:
- ClimbMover.component/init.luau:130-132 declares searchSide, searchTravel and SEARCH_BUDGET (2.0) above edgeUnderBlock (781) and kickOff (742) above it, so declaration order holds.
- edgeUnderBlock with searchTop: 796-804 records the side the search set out on and, when openSide later answers anything else (the other side, or 0) after the slide is blocked, lets go 'no way up' (798-801) instead of turning back. 810-815 adds each step to searchTravel and lets go past SEARCH_BUDGET. A capped post between two higher posts now ends in one slide and a let-go, never to and fro.
- With searchTop false (the under-block call, 1126) nothing changed: the fits(p+0.1) early return (789-795) and the dead-end timer behave as in 1.1.22, and the search state is never set (804/810 are guarded by searchTop). Mantling off still passes false at the blocked top (1137), so it holds still.
- Resets: handBack (433-434), the over-lip end (718-719) and both places slideSide is reset in climbUpdate (1132-1133, 1143-1144), plus the early return in edgeUnderBlock (792-793).
Over-lip duration (932): mantleDuration is bounded by mantleMaxTime and the factor is path / max(rise, 0.5) floored at 1, so there is no divide by zero, and with depth < 0.8 and rise <= hangDepth+0.4 it stays within a few seconds; mT also divides by max(mDuration, 0.05) (697).
Docs: ClimbMover README:63 has onKickOff back in the event list and 'no way up' described; README:65 states the over-lip order the code fires (onMantleStart 942; onMantleEnd then onClimbStart 725-726; from a hang onHangEnd 'over lip' at 1324 after climbOverLip starts, before any end event). README.md:135 (1.1.20 bullet) now mentions the 1.1.22 pinned let-go; 1.1.22 and 1.1.23 changelogs (142, 144) describe the one-way 2 m search and the path-scaled pace.
1.1.22 behaviour intact: pinned let-go (1155-1167), straight-up fallback (1105-1113) and the climbOverLip 0.8 m limit (906-912) are unchanged.
Nit (not blocking): startClimb (483-497) resets deadEndFor and pinnedFor but not slideSide/searchSide/searchTravel. It cannot leave stale state today, since every climb ends through handBack or the over-lip end and tryClimb needs state idle, but resetting them there too would make the invariant local.
Sign in to comment or review.