federicobravetti/arc-agi-3

ci: inject HF Space frontmatter

fix(agents): under fog, explore before chasing transforms (L7 one-life)

fix(perception): stop the L7 goal hitbox inflating into the fog

feat(engine): START_LEVEL env var to begin a local run at a chosen level

refactor(agents): move abstract_agent + my_agent into agents/templates/

fix(server): make /api/stop actually halt the agent, not just the swarm wait

fix(agents): explore/rescue avoid the WHOLE moving-cross line, not just its cell (L7 over-rotation)

perf(agents): lock the moving-cross cycle from fewer fog sightings (L7 one-life, change 1/N)

fix(agents): reject off-line cross detections once its line is known (L7)

fix(agents): stabilise the goal under fog (remember it; box was jittering on L7)

fix(agents): gate the bounce-cycle outlier handling to fog too (L6 first-try)

fix(agents): gate ALL fog-specific logic to fog levels (L6 regression)

fix(agents): outlier-robust bounce cycle -> the moving cross is finally learned (L7)

fix(agents): remember the colour-tile order so the recolour count is exact under fog

fix(agents): know a recolour is needed even when the colour tile is out of view (L7)

fix(agents): forget a transform-tile belief when stepping on it does nothing (L7)

feat(agents): go DO the next needed transform action by chasing its tile (L7)

fix(agents): gap-tolerant cycles for all moving tiles + longer trajectory log

fix(agents): hybrid fog tile-tracking -- clean map + remember each tile's last cell

fix(agents): remember each moving tile's LATEST position (not floor-ify it away)

fix(agents): no ghost moving-tiles in the fog map; learn their motion live (L7)

perf(agents): single BFS for fog frontier exploration (was O(frontiers x map))

perf(planning): memoize budget-planner leg/reach pathfinds (L7 was seconds/frame)

feat(server): viewer shows the agent's learned fog map, not just the visible disk

fix(agents): don't accumulate the live HUD reference into the fog map (L7)

feat(agents): fog-of-war exploration for level 7 (remembered map + frontiers)

fix(agents): make the phase-1 shape-cycling trial energy-aware (g2)

fix(agents): gap-tolerant shape-tile cycle so phase 1 intercepts it (g2)

fix(agents): keep the shape requirement when the shape tile is out of sight (g2 stall)

feat(agents): log per-frame plan decisions to a .plan.jsonl beside the recording

fix(agents): intercept the MOVING shape tile during trial shape-cycling (level 6 goal 2)

fix(agents): remember the occluded cross so the agent never freezes (level 6 first try)

fix(agents): rotation stays possible when the moving cross is occluded

fix(agents): never swap the shape when the goal is the current shape rotated

feat(planning): intercept moving colour & shape tiles (safe re-apply, no observe gate)

fix(ls20): persist game-wide beliefs across levels; detect reference when its box lights white

feat(planning): multi-goal level 6 (reach both goals, one at a time)

feat(perception): detect multiple goals (level 6 has two)

fix(agents): remember the shape tile under the player (occlusion) so shape-cycling keeps swapping

feat(agents): cycle the shape tile by trial (and learn its sequence) for level 5

fix(agents): discover shape swaps by trial, rotate only on a matching rotation

feat(planning): compose shape-swap with rotation for level 5

feat(planning): intercept the level-5 moving cross (predict + classic pathfinder)

feat(agents): learn the level-5 moving-cross pattern (BDI)

fix(planning): model a slide as two moves (step on, then teleport)

feat(agents): learn the level-4 shape-swap cycle (BDI)

feat(planning): plan the level-4 shape swap as a third transform resource

fix(viewer): cleaner slide-direction arrows

feat(viewer): draw the level-4 shape-swap tile (S marker)

feat(planning): detect the level-4 reference (any colour) and shape-swap tile