Skip to content

MEMORY.md

Latest commit

 

History

History
43 lines (42 loc) · 3.54 KB

MEMORY.md

File metadata and controls

43 lines (42 loc) · 3.54 KB

Session Memory

  • Harness mode must not block on project picker when ?harness=true and no ?project= is present. Default to default.ltproj (or first discovered project) for deterministic automated testing.
  • In animation combat, cyan/red solid rectangles indicate the deliberate sprite stub fallback path (draw.mainFrame is missing) in src/engine/states/game-states.ts.
  • Fixed by gating AnimationCombat init: src/combat/animation-combat.ts now waits until both sides resolve mainFrame before entering visible phases (SPRITE_LOAD_WAIT_MS = 1500 fail-safe).
  • Harness loadLevel(clean=false) must not manually push event after free; FreeState already auto-pushes pending level_start events. Double-pushing can stack EventState and cause chapter intro soft-lock/transient empty-top-state behavior in long cutscenes (seen in Ch.2/Ch.3).
  • Talk menu discovery for level-scoped conversations must pass levelNid into eventManager.getEventsForTrigger(). Missing levelNid hides valid Talk options even when matching events exist (reproduced with Ch.5 Natasha/Joshua recruitment).
  • For long village cutscenes in harness tests, use BACK input to enable EventState skip mode; relying on repeated SELECT can leave tests mid-event and make item assertions flaky.
  • evaluateCondition(unit.can_unlock(region)) must handle runtime item components stored as Map<string, any>, not just array-shaped components. Include can_unlock expression handling (e.g. region.nid.startswith('Chest') / 'Door') for correct Chest/Door menu gating.
  • In multi-step interaction tests, prior actions can leave units with finished=true; either reset per-turn flags explicitly or reload the level cleanly between cases to avoid false negatives when forcing menu state.
  • Destructible village data may be mixed between DestroyVillageX interaction regions and event conditions that match VillageX. Region-trigger code should retry with sibling region context when Destructible trigger fails on Destroy* region.
  • Ch.3 Colm spawn event (3_Turn2) uses trigger other_turn_change with condition game.turncount == 1 in this data set; harness tests should set both turnCount and turncount before triggering for deterministic behavior.
  • Ch.3 outro validation should assert recruit branch behavior during the outro timeline (Colm becomes player) rather than after Ch.4 loads, since Ch.4 intro removes many units.
  • Ch.4 Village2 recruitment (change_team;Lute;player + add_unit;...;closest) can leave Lute unplaced (position == null) in some deterministic harness setups; assert team conversion and region consumption as primary correctness checks.
  • Ch.4 Turn3 cameo script intentionally removes L'arachel, Dozla, and Rennac after dialogue; regression checks should assert final position == null rather than persistent on-map presence.
  • ralph-loop only runs when PLAN.md has unchecked - [ ] items; keep backlog items concise so prompt titles stay short and avoid Discord attachment fallback.
  • Magic sword harness regression should assert deterministic execution signals (state resolution, weapon uses decrement) rather than RNG-dependent HP damage.
  • Event dialog parity: portrait mouth animation must follow dialog typing state, not just dialog lifetime. Stop talking when dialog enters waiting/pause and only animate during active typing.
  • Event background parity: change_background should block event progression until panorama load resolves; async non-blocking loads can cause first cutscene lines to render without background.
⚡