Scenario Godot Gameplay

Godot gameplay systems (gameplay systems programmer)

Target: Godot 4.7.2.stable, Jolt, macOS Apple Silicon.

At expert level, gameplay systems are small contracts that can be tested without a scene: pure decision functions, Resources for data, groups for discovery. Their cost is measured at the agent count the game needs before anyone argues about it. The core stance: build the rule as data, run it headless at --fixed-fps 60, then look at one capture. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

REQUIRED BACKGROUND: scenario-godot-expert (channels, gd_run, the review loop, 4.7 traps).

Status (2026-10-02): 16 live tests, the round-2 checks (P15) and the offline suite pass in Godot 4.7.2 (references/procedures.md, tests/code/godot-gameplay/). Dialogue Manager 4.1.0, LimboAI 1.8.1 and GUT 9.7.1 were run in their own project copies.

Stance (the expert delta)

1. Benchmark with --fixed-fps 60, never plain headless. A headless frame sleeps 6.9 ms (low_processor_usage_mode_sleep_usec) and runs fewer physics ticks than frames. Performance.TIME_* monitors refresh about once per second and read 0 under fixed fps. Measure wall milliseconds per frame, median of 3, and compare A/B pairs from the same run (observed, P0).
2. The crowd is not the bottleneck the baseline thinks. 200 CharacterBody3D enemies with NavigationAgent3D avoidance cost about 2 ms per frame on an M5 Max. Cost is linear to 800 agents (7.3 ms), and avoidance itself is within noise at 200. Tuning max_neighbors or the repath interval gave no measurable gain. What avoidance buys is correctness: 0 overlapping pairs against 3108 without it (P2). Go to server agents plus one MultiMesh (0.26 ms for 200) only when the agents need no nodes.
3. Wait for the mesh you assigned, not for "the map". The docs say wait one physics frame. Live, the first sync took 3. After a rebake the map id moved while the region still served the old polygons. Read region_get_iteration_id before assigning, wait for it to move, then one more map sync (nav_tools.wait_applied, P1).
4. Bake from a group of static colliders. A new NavigationMesh parses Both (meshes and colliders) from the region's children, so an NPC under the region carves a 3 m hole (Bramwell 2W4JP48oZ8U 00:12:37; measured P1).
5. Pick the AI tier by the deciding condition. Use an enum brain for crowds (Shaggy Dev oqFbZoA2lnU 00:00:32) and a node FSM for a hero or player character. Use a behavior tree when designers author behavior (Chap.C E_FIy2dTkNc), and GOAP only when many generic actions recombine (Wahaha Yes cm5Jxo31plw). Budget thinking: 5 Hz with 40 enemies per tick cost 0.07 ms over no AI; every enemy every tick cost 0.45 ms (P3).
6. Hitboxes mask, hurtboxes layer, one HitLog per swing (Queble cX-vzfmzjnE 00:06:36, 00:12:55). New areas start on layer 1 and mask 1 and hit their own owner (P4).
7. Fast bullets are data, or bodies with CCD. Without continuous_cd, 6 of 8 balls went through a 0.1 m wall even at 30 m/s. Jolt clamps at 500 m/s. Ray-swept data bullets drawn by one MultiMesh cost 0.35 ms per 1000, against 1.05 ms for pooled RigidBody3D with CCD (P5).
8. A .tres save is code. A GDScript sub_resource runs on load (Godotneers 43BZsLZheA4 01:03:14; reproduced P6). Scan before loading, load with CACHE_MODE_IGNORE, version and migrate, validate values instead of encrypting (01:01:37).

The NavigationAgent3D contract (P15, measured)

Docs warnings (NavigationAgents page) and Bramwell (2W4JP48oZ8U 00:14:14), measured in 4.7.2 unless marked:

• Avoidance needs a target_position. With avoidance on and no target, safe_velocity stayed 0 for 30 ticks; with a target it returned the full 3 m/s.
• Avoidance never feeds pathfinding. Agents are circles with no navmesh or physics knowledge, so walls still come from the path (docs).
• One get_next_path_position() per physics frame, and none after the end. Once is_navigation_finished() is true it still returns the last point (0.93 m away here): stop steering there, or the agent creeps.
• path_desired_distance must exceed one tick of travel and the navmesh's float above the floor (about 0.3 m). At 25 m/s (0.42 m per tick) with 0.1, the agent reversed on 1782 of 1800 ticks and never arrived; at 1.0 it arrived in 182. With 0.1 and the feet measured at floor height, it never left the first path point.
• Defaults: time_horizon_agents 1.0, time_horizon_obstacles 0.0. Lower the horizon if agents slow down too early. Too few max_neighbors and avoidance is ignored (docs; not measured here).

Establish first

Workflow

1. Set up the project.
   • Clone Base3D with gd_env.base_project("3d", P) and give it its own application/config/name, because clones otherwise share user://. Run gg.install(P).
   • GATE: project.godot has 3d/physics_engine="Jolt Physics" and a stated renderer; gd_run.check_all(P) reports 0 parse errors.
2. Level navigation.
   • Static geometry goes in group nav_source. Build the mesh with make_navmesh(parsed colliders, source group), then bake, then wait_applied.
   • GATE: map_get_path between sample points is non-empty with 0 obstacle crossings; the polygon count is above 0; no hole at NPC spawns (P1).
3. Movement and the crowd.
   • Enable avoidance; velocity_computed drives move_and_slide; repath is throttled; steering is XZ only.
   • Run gg.crowd_bench(P, n) at 50, 200 and 800, then gg.scaling_verdict. In your own benchmark scene, log Performance.add_custom_monitor(&"ai/thinks", callable) and OBJECT_ORPHAN_NODE_COUNT beside wall time (it read 5 after 5 stray Node.new(), P15).
   • GATE: 0 overlapping pairs; exponent under 1.3; the median at the target count fits the gameplay budget (gg.budget_share); the orphan count is flat. Then a windowed capture of the crowd, opened.
4. AI.
   • Write the brain as decide(state, percept) with hysteresis. Perception is cheap first (distance, cone, one ray, stale-hit check). The manager thinks at about 5 Hz with a per-tick cap. Far enemies think less: key the tiers on squared distance, because VisibleOnScreenNotifier3D.is_on_screen() stays false headless and a benchmark would put everyone in the far tier (P15). The node FSM is for the hero; LimboAI if designers author.
   • GATE: the GUT transition table is green; the hero scenario history matches the expected sequence (patrol, chase, attack, chase, search, patrol); the think-cost A/B is measured (P3, P9, P13).
5. Combat.
   • Name the layers (World 1, Player 2, Enemy 3, PlayerHurtbox 5, EnemyHurtbox 6, Projectile 7). Hitbox and hurtbox come from the kit. Bullets go through bullet_pool; bouncing objects are RigidBody3D with CCD.
   • GATE: a shared-log swing scores 1 hit; same faction 0; the tunnelling test reads 0 of 8 at the top speed; bullets on their own layer (P4, P5).
6. Persistence.
   • Typed save Resource with VERSION and migrate(), save_io.save_resource_atomic, load_resource_checked; inventories duplicated per holder.
   • GATE: round trip equal; every old fixture loads; the evil fixture is refused; the cache test passes (P6, P7).
7. Dialogue and quests.
   • Dialogue Manager 4.x .dialogue files; quest order is turn in, then continue, then offer.
   • GATE: the import compiles every file with 0 errors; gg.check_dialogue_text finds no missing cue; a headless walk reaches each branch and mutates state (P8).
8. Input.
   • Physical-key actions; gameplay in _unhandled_input; remaps go to a ConfigFile.
   • GATE: synthetic press and just-pressed tests pass; the STOP panel consumes a pushed click; the remap survives a reload (P10).
9. Procedural content.
   • Seeded data first (dungeon_gen), validators, then geometry, bake and path.
   • GATE: 1000 seeds with 0 validator failures; the same seed gives the same hash; a path from start to boss exists; the top capture is opened (P11).
10. Wrap up.
    • Run test_gameplay_live.py and copy the evidence. Score yourself against references/critique.md.