---

name: scenario-godot-gameplay description: "Use when building gameplay systems in Godot 4.7: enemy AI that patrols, chases and attacks, state machines, behavior trees (LimboAI), NavigationAgent3D pathfinding and avoidance for 100 to 1000 enemies, 'agents stack up' or 'path is empty', hitbox and hurtbox combat, fast bullets going through walls, save and load, inventory, quests, dialogue (Dialogue Manager), input remapping, procedural dungeons, or measuring agent counts and frame time headless." license: MIT

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

Input	Why it changes the plan	Default
Agent count on screen	nodes vs server agents, think budget	200
Frame budget for gameplay	the gate for every benchmark	a quarter of 16.67 ms (4.2 ms) at 60 fps [added]
2D or 3D	NavigationAgent2D vs 3D, Area2D vs 3D	3D
Who authors behavior	code (FSM, enum) vs designers (LimboAI)	code
Save trust	own files only vs shared or modded saves	shared: scan or JSON
Platform	input devices, CPU budget	desktop, keyboard and mouse plus pad

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.

Numbers

Value	Measured	Relative to
200 bodies with avoidance	1.98 ms per frame, 0 overlaps	M5 Max, fixed fps 60, 60 m arena with 14 obstacles
Scaling with avoidance	50 agents: 1.07 ms, 200: 2.00, 800: 7.31 (about 9 µs per agent)	same
200 server agents plus MultiMesh	0.26 ms	same
Think cost, 200 enemies	budget 0.40 ms, every tick 0.78, none 0.33	5 Hz, 40 per tick
First navmesh sync	3 physics frames; rebake applied after 2 to 16	async iterations on (default)
Bake	5 ms, 180 polygons	60 m arena, cell 0.25
1000 bullets	data 0.35 ms; RigidBody3D with CCD 1.05; without CCD 1.00 (45% miss walls)	60 m/s, 2 s life
Jolt speed clamp	500 m/s	physics/jolt_physics_3d/limits/max_linear_velocity
Dungeon generate plus validate	88 µs median, 1000 of 1000 valid	14-room main path, 4 branches
NavigationAgent3D defaults	radius 0.5, max_neighbors 10, neighbor_distance 50, max_speed 10, avoidance off	4.7.2 ClassDB

Quality gates

• Measurable: crowd metrics (overlapping pairs, min pair distance, stuck, fell through the floor); scaling_verdict; budget_share; GUT (gd_run.run_tests) green after one deliberate red; check_all with 0 parse errors; validators on 1000 seeds; save fixtures; the dialogue compile.
• Visual: windowed captures in a SubViewport (160x90 window), with gd_review.image_checks and a contact sheet of 1600 px or less, opened and looked at. Look for crowd spacing round the player, no agents inside obstacles, and the dungeon path from start to boss.

Common mistakes

Mistake	What it looks like	Fix
Benchmarking plain headless	about 7 ms per "frame" whatever the load	--fixed-fps 60
Path queried right after assigning the mesh	empty path, agents idle at spawn	wait_applied
Default bake (Both from root) with NPCs inside	holes round spawn points	group plus static colliders
Avoidance off "for speed"	3108 overlapping pairs, enemies merge	avoidance on: it costs little at 200
Vision as one Area3D per enemy	broadphase churn	distance, cone, one ray [added]
Trusting a ray hit on a teleported body	enemy "sees" through a wall	stale-hit distance check
Hit areas on layer 1 and mask 1	self-hits, triple damage	mask or layer only; HitLog
monitoring = false in area_entered	"Function blocked during in/out signal"	set_deferred
Bullet bodies on the default layer	1000 bodies collide: 110 to 170 ms per frame	own layer, mask world and targets
No CCD on fast bodies	bullets pass walls at 30 m/s	continuous_cd or ray-swept data
Loading a downloaded .tres	its script runs	load_resource_checked or JSON with from_native
Default load() after re-saving	old values come back	CACHE_MODE_IGNORE
save.items = [...] on Array[String]	runtime error or a silent no-op through set()	items.assign([...])
Preloaded inventory .tres on every chest	one chest's loot shows in all	duplicate(true) per holder
Always-succeeding first child in a Selector	the attack branch never runs (Chap.C E_FIy2dTkNc 00:17:27)	gate first, BTProbability or AlwaysFail
Logical keycodes in actions	WASD breaks on AZERTY	physical keycodes
Teleport with physics interpolation on	the body is drawn sliding across the jump [added, not measured]	reset_physics_interpolation() after the move (method exists)
queue_free() on an autoload	no crash in 4.7.2; every later access is a "previously freed" error reading 0 (docs say crash)	never free one; reset its state instead

Handoffs

• Receives: project layout, autoload policy and save model (scenario-godot-architecture); levels with static colliders in nav_source and the controller (scenario-godot-3d-world); TileMapLayer navigation (scenario-godot-2d); art from scenario-* skills through those two.
• Delivers: state_changed to scenario-godot-animation; play events and priorities to scenario-godot-audio (mix, buses, attenuation and the listener stay there); inventory signals, dialogue balloon and remap menu to scenario-godot-ui; hit position and normal to scenario-godot-vfx; Resources and pure decide() to scenario-godot-multiplayer; benchmark JSON to scenario-godot-performance-export.

Godot 4.7 notes

• Navigation classes are experimental in 4.7.2. Region and map iterations are async by default, so syncs take several frames. region_bake_navigation_mesh is deprecated: use parse_source_geometry_data plus bake_from_source_geometry_data (or the async variant). region_get_iteration_id exists. query_path(params, result, callback) ran its callback before returning (P15): it does not move work off the frame.
• A MultiMesh culls as one box: set custom_aabb (exists) and split big crowds by area.
• The NavigationMesh default geometry_parsed_geometry_type is Both (2). Default cell size and height are 0.25.
• Jolt is the 3D default for new projects from 4.6, but an agent-written project.godot gets GodotPhysics3D unless it sets Jolt (lead's delta).
• duplicate(true) copies only file-internal subresources (4.5 and later). duplicate_deep(DEEP_DUPLICATE_ALL) copies external ones too.
• JSON.from_native and to_native exist; plain JSON numbers come back as floats.
• InputEventKey.as_text() prints "W - Physical". A new InputEventKey has device 16 (DEVICE_ID_KEYBOARD).
• Addons: Dialogue Manager is at 4.1.0, where titles are cues; DM3 videos are a version behind. LimboAI 1.8.1 ships a "4.6" GDExtension build that runs on 4.7.2.

References

• references/procedures.md (P0 to P15 with code and live results), expert-notes.md (judgment by expert, timestamps), critique.md (rubric), gui-paths.md (editor menus), sources.md (videos, docs, addon hashes).
• scripts/gd_gameplay.py: install, run_job, crowd_bench, scaling_verdict, budget_share, dialogue lint, save scan.
• scripts/agentkit/gameplay/: nav_tools, arena, crowd_bench, enemy_body, state machine and states, brain, perception, enemy_manager, hitbox, hurtbox, hit_log, combat_layers, bullet_pool, save_io, inventory, quest_rules, dungeon_gen, voice_pool.