Manager scripts driving many objects from one system

The recipe

Author with project_apply_script_semantic_patch using the ai_safe op pack. The manager is an entity with a ScriptComponent.

1. Tag the cohort: give every member a shared tag (for example mob) so one query finds them all.
2. Enumerate on a beat: addTickHandler { "handlerKey":"onManage" } (or a coarser addTimerHandler for slow updates). Inside, findEntitiesByTag { "tag":"mob", "assignTo":"mobs", "limit": 64 }, or use the paged ctx.world.query in a GessaScript body for larger sets (pass its nextCursor back verbatim to page).
3. Act on each member: forEachEntityInResult { "arrayRef":"mobs" } whose body uses only the loop-intent ops (despawnEntity, emitCustomEvent, incrementNumericState, attachComponentWithValidatedDefaults). For richer per-member work (move, retarget), iterate in a GessaScript body and call ctx.entity.patchComponent on each.
4. Spawn and cull centrally: keep the population at a target size by spawnEntityFromTemplate when short and despawnEntity the excess or the dead, all from the one manager.
5. Shared state: hold cohort-level counters (alive count, wave number) in durable state with incrementNumericState so the manager and the HUD read one truth.

Pitfalls

• forEachEntityInResult.body accepts ONLY the four loop-intent ops; anything richer must run in a replaceScriptFromGessaScript body iterating the result.
• Bound every scan: cap findEntitiesByTag limit and forEachEntityInResult maxIterations (both cap at 1000) and page ctx.world.query rather than pulling everything each tick.
• The collision and per-tick budgets cap at 1000 entities; a manager that keeps spawning without culling will hit the ceiling.
• Run the manager on the authority; do not also run a client copy or the cohort double-updates.

Verify

• simulation_run (qa.run.start) to watch the manager spawn, update, and cull the cohort over several ticks.
• project_get_graph_snapshot to confirm the single manager ScriptComponent and its tick or timer handler.