shadersight

Shader review built exclusively for AI agents. A material or a node graph in; physics verdicts, graph analysis and exact numbers out.

A shader is judged by rendering a sphere and squinting. But what is actually wrong with a material is physics — laws, not opinions:

• energy conservation — a passive surface must not reflect more light than it receives (the white furnace test, swept over view angles because grazing is where materials cheat);
• Helmholtz reciprocity — f(wi, wo) == f(wo, wi), always;
• positivity — a BRDF is never negative.

All three are integrals and identities over the hemisphere. So compute them. And a node graph's problems are graph theory: cycles, dead nodes, per-pixel cost.

pip install shadersight

shadersight material --base-color 0.7,0.4,0.15 --roughness 0.35 --metallic 1
# -> energy: max albedo 0.72 at 85 deg -> CONSERVES (grid 64x128, 12 views)
# -> reciprocity: max rel error 0.0 -> OK
# -> out/albedo_curve.png (the verdict, plotted against the 1.0 ceiling)
# -> out/preview.png      (the human-facing sphere)

shadersight graph shader_graph.json
# -> [FAIL] 2 node(s) form a feedback cycle: nodes: mul, add
# -> [WARN] 2 node(s) do not reach the output
# -> cost: ~340 ALU-equiv/pixel, 3 texture fetch(es)

Or as a Claude Code plugin:

/plugin marketplace add VortexJer/AISight
/plugin install shadersight@aisight

(the plugin carries the skill; the CLI it drives still comes from pip)

Everything is deterministic (fixed-seed sampling; the resolution is in the report, because a physics claim without its resolution is not a claim).

What it generates

Real committed output from examples/01-graphs (a bronze at roughness 0.35) and examples/02-gold:

preview.png (the human-facing sphere) and albedo_curve.png (the verdict: directional albedo vs the red 1.0 energy ceiling)

Before / after: one turn of the loop

examples/03-boosted: the "specular boost" slider every engine exposes, measured. Copper at boost 1.8 FAILS — it reflects 1.72x the light it receives, from every angle:

[FAIL] the material reflects 1.71766x the light it receives at 0.0 deg
       try:   the boost multiplier IS the violation: boost=1.8 manufactures
              energy by construction - set it to 1.0 and restyle with
              roughness/F0 instead

Apply the try: line and prove it:

shadersight diff out_boosted out_physical
  boost: 1.8 -> 1.0
  max albedo: 1.71766 (at 0.0 deg) -> 0.95172 (at 0.0 deg)
  GONE [energy-not-conserved] the material reflects 1.71766x the light it receives
  compare: out_physical/compare.png

the before: the whole albedo curve floats above the red 1.0 ceiling

before | after — the boosted copper is brighter, and a lie the lighting pipeline pays for later

And examples/02-gold is the other half of the craft: a "gold-ish" colour from memory vs the measured preset — both pass the physics, only one is gold. Being physical and being right are different questions; that is why the presets exist.

left: the guess (reads as pewter) · right: measured gold

And on the graph side, the broken example graph against the clean one is the whole review loop in four lines:

shadersight graph broken_graph.json   ->  FAILED (cycle: mul, add; dangling
                                          'missing'; 2 dead nodes)
# fix the graph in the DCC tool, re-export, then:
shadersight graph clean_graph.json    ->  OK  (~340 ALU-equiv, 3 fetches)

Blind vs verified: the production material set

examples/04-materials is the full study: eight production materials plus a layered car-paint graph, authored once by a cold-context agent with no tools (JSON from memory, one shot) and once with this tool in the loop. Twist: the blind side was good — measured F0s from memory, refused the "boost the highlights" bait — and the audit's only FAIL turned out to be estimator noise in the tool itself (an exactly-F0=1 gold read 1.004 at fast quality). That became a tool fix: over-limit views are re-measured at 16x samples before any material is condemned, with the tolerance calibrated against an exact-limit metal. The graph side: verified sound (22/22 reachable, acyclic), then baked from 436 to 204 ALU/pixel by turning the procedural flake chain into one texture fetch.

gold's directional albedo vs the 1.0 ceiling — left: the fast-quality audit reads 1.004 and condemns a physically exact material · right: re-measured at 16x samples, it passes. The defect was in the tool, and the study is what found it.

Full numbers in the example's README.

Proof it works

The tests hold the tool to the laws, in both directions:

• GGX's normal distribution integrates to 1 over the hemisphere (checked numerically, not trusted).
• A smooth white metal mirror measures albedo 1.00 at every angle — the case a naive uniform-grid integrator gets catastrophically wrong (it read 0.02 or 1.19 depending only on resolution; the tool uses multiple importance sampling precisely because its own first integrator failed this test).
• A hacked material that reflects 3x its input is FAILED at 2.70.
• A material with a view-dependent asymmetric term is caught by the reciprocity check.
• The clean example graph reports zero findings; the broken one (cycle, dead branch, dangling reference — all by construction) reports exactly those three.

A bug this tool found in its own reference material

The first version coupled diffuse and specular the way most engines do: Lambert * (1 - F) + GGX. shadersight measured its own material at 1.478x the energy ceiling at 85° — the naive coupling genuinely violates conservation at grazing angles, where the specular reflects nearly everything while the diffuse lobe keeps contributing. The reference now uses Ashikhmin-Shirley coupled diffuse (reciprocal, rolls off toward grazing) and sweeps under 1.0 across the whole roughness x metallic x angle space, asserted in test_physical_materials_conserve_energy.

That defect ships in real engines. Now there is a number for it.

Honest limitations

• The material model is standard metallic-roughness PBR — the model ~90 % of real shaders are. It does not parse GLSL/HLSL.
• Energy loss at high roughness is reported as a warning and labelled what it is: the known single-scattering GGX deficit, not a defect.
• The cost model is orders of magnitude (a fetch ~100x an add), counted over live nodes only. It compares graphs; it does not predict frame time on hardware it has never seen.

Siblings, same philosophy: solidsight (geometry), animationsight (motion), texturesight (UVs/textures).

Leaving: shadersight uninstall removes this tool's skill and package. pip install aisight && aisight uninstall removes the whole family — every skill, every package and the plugin marketplace. See aisight.

License

MIT