ShaderToy CLI agent workflow
To start from an existing public ShaderToy:
shadertoy import https://www.shadertoy.com/view/XXXXXX -o project
This creates the same editable directory format used by shadertoy new and keeps
supported remote assets local.
1. Start by inspecting the project:
shadertoy inspect --json
2. Edit ShaderToy.toml and files under shaders/ or assets/.
Put quality tiers in `[preset.NAME]` / `[preset.NAME.pass.PASS]` and select
them with `--preset NAME` instead of copying the project.
For GPU simulations, FFTs, particles, and large structured state, prefer
compute passes + typed targets + named SSBOs; see shadertoy docs passes.
The manifest is schema-backed. To print the exact schema:
shadertoy docs manifest --schema
3. Validate after edits:
shadertoy check --json
For advisory graph/resource linting too:
shadertoy check --pedantic --json
shadertoy graph --dot target/graph.dot --json
Or validate an effective quality tier:
shadertoy check --preset medium --json
4. Render deterministic evidence:
shadertoy render -o target/check.png
For temporal comparison/contact sheets, reuse one runtime:
shadertoy render-frames --range 0:180:60 --contact-sheet target/contact.png
For an encoded deterministic clip:
shadertoy render-video --frames 180 -o target/clip.mp4
For parameter/art-direction comparisons:
shadertoy sweep --frame 120 --set foam_gain=0.8,1.0,1.2
To avoid value/expectation bias, prefer a blind comparison when choosing a visual variant:
shadertoy sweep --blind --frame 120 --set foam_gain=0.8,1.0,1.2
To blind old-vs-new images, projects, builds, or git revisions:
shadertoy blind create target/old-renders target/new-renders --output-dir target/old-vs-new
To compare quality presets without temporary project copies:
shadertoy blind create 'project:.@preset=high' 'project:.@preset=medium' 'project:.@preset=low' --frames 60,180,300 --set u_storm=1
shadertoy blind judge target/old-vs-new/blind-session.json --pick B --reason "concise visual rationale"
shadertoy blind reveal target/old-vs-new/blind-session.json
For a complete baseline/candidate experiment with metrics, provenance,
profiling, contact sheets, and optional blind labels:
shadertoy experiment --baseline git:main --candidate git:HEAD --frames 0,60,120 --set u_storm=1 --blind
5. Debug multipass projects from the outside in:
shadertoy inspect graph --json
shadertoy inspect pass buffer-a --json
shadertoy inspect buffer buffer-a --frame 120 --pixel 8,8 --json
shadertoy inspect buffer gbuffer --output-index 1 --raw target/attachment.rgba32f
shadertoy inspect storage particle-state --frame 120 --type f32 --count 16 --json
shadertoy render --pass buffer-a -o target/buffer-a.png
6. Freeze a feedback state when a bug appears:
shadertoy state capture --frame 300 --include-storage --set storm=1.0 -o target/frame300.ststate
shadertoy state inspect target/frame300.ststate --json
7. Replace a buffer with a known image to isolate a pass:
shadertoy render --state target/frame300.ststate \
--set-buffer buffer-a=fixtures/a.png \
-o target/debug.png
8. Profile expensive passes on the real GPU path:
shadertoy profile --frame 120 --warmup 5 --samples 30 --json
Never claim a performance improvement from lowering the final Image output
resolution. With `--preset NAME`, profile locks the benchmark to the base
project's output dimensions by default; preset `render_scale` is reported
separately and excluded from the comparison. Optimize internal pass
dimensions, iterations, algorithms, or shader work instead.
Profiling separates the raw GPU execution timestamp, CPU completion-wait
duration, and legacy post-completion attributed interval for every pass/sample.
Async compute timers that outrun the actual work are marked sample_valid=false
instead of being presented as trustworthy shader cost. The report also includes
an independent non-intrusive frame-level GPU timestamp interval and MAD-based
outlier flags. Prefer gpu_frame_timestamp for cross-run optimization comparisons;
use per-pass validity/wait data to localize changes.
Use --discard-outliers to exclude flagged samples from aggregates while keeping
all raw sample_details. For maximum-isolation diagnostics, rerun with:
shadertoy profile --frame 120 --samples 30 --sync-per-pass --json
See shadertoy docs profile for interpretation guidance.
9. Define deterministic [[test]] cases in ShaderToy.toml and run:
shadertoy test --ci
Tests can cover frame/resolution matrices, uniform-vs-uniform RMSE bounds,
per-pass/total GPU budgets, exact SSBO fixtures, deterministic raw output,
output-resolution independence, and .ststate round-trips. Use --update
deliberately to write visual baselines. See shadertoy docs test.
10. Freeze difficult deterministic renderer bugs into a self-contained bundle:
shadertoy trace capture --frame 120 --include-intermediates -o target/bug.sttrace
shadertoy trace inspect target/bug.sttrace --json
shadertoy trace replay target/bug.sttrace --json
The trace carries STTF, persistent state/SSBOs, timings, hashes, graph
diagnostics, and optional pass readbacks. See shadertoy docs trace.
11. Use live native-rendered review when iterating:
shadertoy preview
Or launch a named quality tier directly:
shadertoy preview --preset medium
When presets exist, the browser also exposes a Quality preset selector for
live switching. Final Image output stays fixed at the project review
resolution across preset changes so a smaller image cannot masquerade as a
performance win. The human can select common review resolutions or enter a
custom size; that explicit size also stays fixed across presets.
Declared custom uniforms become live controls. A `kind = "webcam"` channel
exposes a Start webcam button; webcam input is intentionally not recordable
or usable by headless commands.
Shared GLSL can use quoted #include directives; preview recompiles only passes
affected by a changed source/include and keeps existing feedback targets alive.
12. Record an input-sensitive preview bug, then reproduce it without the browser:
shadertoy preview --record target/repro.strec
shadertoy replay target/repro.strec -o target/replayed.png
On Linux, check/render/render-frames/state capture/preview/profile/test/experiment/trace/replay use surfaceless EGL and do not require DISPLAY or WAYLAND_DISPLAY. The context prefers OpenGL 4.3 and falls back to 4.1 for projects that do not use compute/SSBO features.
Do not edit target/. It is disposable generated output.