Scenario Godot Shaders

Godot shaders (technical artist, shading)

Target: Godot 4.7.2.stable, macOS Apple Silicon (Metal by default, Vulkan through MoltenVK for GPU timing).

Expert level here means every shader is text the agent can diff, compiled under each target renderer, drawn on a real mesh in a window, swept across its parameters and looked at on a contact sheet, with its GPU cost measured against a StandardMaterial3D baseline. Compiling is not the same as working: several 4.7.2 failures only appear when the backend draws, and several bugs produce no error at all. 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).

Stance (the expert delta)

1. Code first, graphs on request. Text shaders give loops, functions, #include and clean diffs (Godotneers, Shaderland [01:10:09]; Daniel Ilett on the gaps of VisualShader). When a human wants a graph, build the VisualShader in code (add_node, connect_nodes, an Expression node for the maths) and save it as .tres; port names are not exposed to GDScript, so output indices are hard-coded and checked by reading get_code() [added, run 2026-10-02].
2. Compile headless, then draw windowed. Headless get_shader_uniform_list() catches language errors per --rendering-method in about 0.4 s per renderer. Stencil reads in the opaque pass and fog shaders under Compatibility only fail when a window draws them. Uninitialized locals and DEPTH written in one branch never fail at all, so the linter checks those [added, run 2026-10-02].
3. World or object space beats mesh UVs. Dissolve, wind, water and triplanar effects should be driven from object-space or world-space coordinates, so seams and UV stretch disappear (StayAtHomeDev dissolve [00:09:59], DevPoodle triplanar, Dylearn grass). For a shared material, put the per-object values (bounds, amount) in instance uniforms, or each mesh dissolves at the wrong height [added].
4. Water that does not slide. Use two panned copies of one seamless cellular noise, multiply them, raise to 1.25, and share them between vertex() and fragment() so the crests sit on the wave tops (Bramwell [00:11:57 to 00:20:57]). Use hybrid toon bands, a smoothstep only near each cut, so the bands do not flicker (Dylearn [00:13:42]). Shore foam comes from vertical depth, reconstructed from the depth texture with the reversed-z formula [added].
5. HDR edges need three things together: a color above 1.0, the EMISSION output, and Glow in the WorldEnvironment (StayAtHomeDev [00:07:36], Daniel Ilett [00:10:22], Pixezy [00:06:55]).
6. Know what moves a material to the transparent pass. Writing ALPHA or reading the screen or depth texture moves it there. It loses shadows and sorting, and it is invisible to other screen-reading materials (docs). Use ALPHA_SCISSOR_THRESHOLD for dissolves: the material stays opaque and the shadow dissolves too. Scissor is still an alpha test: like discard, it gives up early-Z for that mesh (docs), which costs most on mobile tile GPUs, so swap the dissolve material in only for the effect (P18). Stencil readers must be transparent, with a render_priority above the writer's (LowQualityCoding [00:07:35]).
7. Compute: on screen, stay on the main device. Texture2DRD needs the main RenderingDevice inside RenderingServer.call_on_render_thread, and no submit or sync (The Pathfinders Codex [00:03:07]). To get results back on the CPU, use a local device, free every RID and test parity against a CPU version (DevPoodle [00:02:21], Crigz [00:11:24]).
8. Bake shaders from a windowed export. The shader baker runs only when the export process has a rendering device. A --headless export with shader_baker/enabled=true silently packs nothing extra [added, run 2026-10-02].

Establish first

• Target renderers: Forward+ (desktop), Mobile (phones), Compatibility (web, old GPUs); each decides which hints and stages exist. Default: Forward+ plus one named fallback.
• Style: toon, PBR or pixel art. Decide the band count, the outline method (stencil or inverted hull) and whether Glow is on.
• Budget: GPU ms per effect against a StandardMaterial3D baseline. Default: under +0.5 ms at 1080p on this Mac for a hero effect, confirmed on the device.
• Who edits: an agent (text) or a human artist too (VisualShader .tres, grouped uniforms with hints).
• Shared or unique materials: per-instance values decide between instance uniforms (at most 16, no textures), unique materials and globals.

Workflow

1. Install and lint. gd_shaders.install(project) copies the templates into res://shaders/ and the tools into addons/agentkit/shaders/. Then gd_shaders.lint(path, renderers, project=P). GATE: no error findings; every warn fixed or justified.
2. Compile per renderer. gd_shaders.compile_matrix(P, root="res://shaders/"). GATE: ok with an empty failed list for every target renderer.
3. Draw once per renderer. gd_shaders.draw_check(P, files) puts each shader on a drawn mesh, canvas item, sky, particle system or fog volume. GATE: zero shader_errors and engine_errors in each run.
4. Build the test scene in code. Write a job that saves a .tscn with the camera, the light and the Environment set explicitly: AgX, Glow on or off as the brief says. Water needs ground crossing the water level and objects above and below it. A dissolve needs at least three meshes with different sizes and rotations. GATE: the scene loads headless with ok, and the project.godot renderer and Jolt settings were checked.
5. Sweep and look. Run gd_shaders.param_sweep(P, scene, shots=...) with kinds shader, instance, global or prop. Open the contact sheet. GATE: changes behave as expected (a monotonic dissolve, a moving phase), and the sheet shows the intended look in every shot.
6. Compare renderers. gd_shaders.renderer_captures(P, scene). GATE: the look holds under each target renderer; known differences are written down (Compatibility renders brighter and more saturated with the same Environment).
7. Cost. gd_shaders.cost_compare(P, {"standard": ..., "hq": ..., "lite": ...}) on Vulkan at 1080p. GATE: the delta against the baseline is within budget; mobile numbers are marked "desktop proxy".
8. Ship. Add a macOS or target preset with shader_baker/enabled=true. Export once from a windowed process (see procedures.md P14). GATE: the log shows "Baking shaders (N steps)" and the pack grows. A missing global uniform shows up here as an error.

Side paths:

• VisualShader for a human artist (P9): build the graph in code with an Expression node and save it as .tres. Read it back with shader_tools.gd:visual_shader, which returns the generated code, the node counts and the uniforms. Then sweep it like a text shader. GATE: the generated code contains every expected output assignment.
• Compute (P12, P13): run compute_probe windowed for the CPU parity check. Use the Texture2DRD job for on-screen output. GATE: zero mismatches, no leaked RIDs, and frames that change.
• Globals (P11): register the global with register_global (it writes the editor's multi-line format). Set live values with RenderingServer.global_shader_parameter_set and keep a copy in an autoload: the getter stalls the render thread (docs). ProjectSettings.set_setting at runtime does not register a global.
• Animating parameters: a Tween or an AnimationPlayer track on instance_shader_parameters/dissolve_amount drives an instance uniform; shader_parameter/<name> drives a material uniform (P8).

Expert checklist (state these in the plan)

• Complete code in the deliverable. Paste the full templates (dissolve.gdshader, water_stylized.gdshader plus its include) and setup script: readers may lack the kit.
• source_color on every color uniform and color sampler, never on normal, roughness or mask data (lint rule source_color).
• Uniform budget 16,384 B on mobile, vec3 padded to 16 B; big arrays go in a texture (lint rule uniform_budget).
• TIME wraps at 3,600 s and keeps running while paused (docs): drive pausable motion from global uniform float game_time, written by the game_time.gd autoload (P18).
• Dissolve bounds. World space breaks when the mesh moves or scales; StayAtHomeDev feeds AABB min and max from an @tool script (pIwQ 09:59 to 12:52). The template uses object space plus an instance uniform vec2 bounds per mesh, which survives movement, scale and sharing.
• Swap in for the effect. dissolve_swap.gd:swap_in(mi, mat) puts one duplicate per surface (keeping each surface's albedo color and texture) and returns the originals for restore (P18).
• Sorting. depth_prepass_alpha with cull_disabled fixes transparent sorting artifacts and keeps shadows (pIwQ 00:01:00).
• LIGHT_VERTEX (4.3) moves the lit position, for example snapped to a grid for stepped pixel-art light, without touching VERTEX (Dylearn 00:01:25); it draws on all three renderers (P18).
• Outlines. BaseMaterial3D stencil_mode outline on dense meshes; an inverted-hull next pass elsewhere (Daniel Ilett 15:18).