Skip to content

armgit / kiln

Public Imported from GitHub
0

Build and revise procedural 3D assets with your coding agent. Local MCP server, CLI and TypeScript engine with rendered review, editable source and GLB export.

8.1 KB
Raw Blame

Kiln scenes

The maintained Farm, Golden Gate, Foundry Floor, Troy, Cabin in the Woods: PSX Horror and shared scene-kit sources live here. Farm, Golden Gate, Foundry Floor and Cabin in the Woods use React Three Fiber; Troy uses Three.js directly. All use the same pinned Three.js version. The original three scenes were integrated from the separate authored scene workspace for the v0.9 site rollout. Generated capture evidence and asset payloads are kept out of Git. Historical milestone documents describe their original runs, not current release acceptance.

Read the source and development instructions in Farm, Golden Gate, Foundry Floor and Troy. Pack source covers Farm, Vehicles, Bridge, Foundry and Troy. Scene controllers and composition remain separate from standalone authored asset programs.

Cabin in the Woods: PSX Horror is a desktop WebGPU asset demonstration under owner review. Its source-only tests run in the portable gate. Asset payloads come from an explicit sealed scene archive; dependency installation and tests do not download them. Its pack, individual gallery assets and scene are staged together for feedback before publication.

Troy's maintained modules, portable tests, bounded asset transport and bundling tools are public source. Its runtime pin selects an optional verified Cloudflare archive: npm run hydrate in packages/troy restores generated models and banks into ignored .runtime/, and npm run dev previews editable source. It requires neither the original author workspace nor a .kiln setup. The normal portable gate includes Troy's source-only Node tests; its generated payloads are never downloaded automatically by dependency install or tests. See its README for clean-checkout commands and production staging.

The backlog tracks remaining engine and scene work. The October alignment plan and milestone ledgers are historical context; Troy source is now included here.

Install with bun install --frozen-lockfile in this directory. Scene dependencies are separate from engine and site dependencies. Run bun run check:toolchain and bun run check:pins first; the pin check resolves ../package.json, ../render-service/package.json, ../site/package.json and ../site/bun.lock within this integrated checkout, and compares the root and scene toolchain pins. Run bun run typecheck, bun run lint and bun run test for the portable source and script contracts; the normal test command now includes scripts/tests. The portable runner explicitly lists 19 author-input suites that belong to bun run test:integration; several other asset-dependent cases also report their existing fixture skips. Browser and asset-integration scripts require staged inputs; authoring scripts that regenerate packs also require their original author workspace. They are not generic clean-checkout tests.

For the production site, follow the deployment guide. site/scripts/scene-inputs.mjs restores the exact reviewed scene packs and standalone bundles under .cache/site-inputs/, with public archive and per-file hash checks. The site builds Farm directly from this source, and serves the reviewed Golden Gate and Foundry chunks without rewriting them. Use the scene kit's sceneStandaloneConfig and verified asset packs when building a new standalone revision; changing source requires requalification and new runtime pins before a site release. After restoring those exact inputs, bun run test:release-inputs exercises the released-asset integration cases and fails when required fixtures are absent.

Current scene inputs are Farm r36-local-review, Golden Gate g9 and Foundry ff3-review2. Those revision names are retained for provenance. Foundry uses AMRs and arms to transport containers and humanoids for maintenance. Its old in-production preview label is historical; current source changes still require qualification and a newly sealed runtime before being served by the site.

Browser qualification on this Linux hub

Scene-kit browser tools and the site's scene island/width/touch/Foundry-flow checks accept host-only configuration. The default remains headless. On the audited NVIDIA Linux hub, use KILN_SCENE_HEADED=1 and KILN_SCENE_CHROME_ARGS='["--ozone-platform=x11","--enable-features=Vulkan","--enable-unsafe-webgpu"]' for hardware qualification. Arguments are a JSON array, not a shell command. An explicit helper headless option wins; helper arguments append after host flags. The 3 October capability probe found SwiftShader with headless Chrome 150 and real NVIDIA WebGPU/WebGL2 with this headed configuration. Record the actual adapter, backend and source receipt for each run; a successful API check alone is not hardware performance evidence. These options affect QA tools, not deployed scenes.

Farm release audit

scripts/capture-farm-paired.ts implements D-68: capture the candidate and an explicit last-accepted test build against the same immutable pilot pair at the same phase. Supply --label, --accepted-root and --new-root; defaults cover all 16 named/play views on both backends. A new failure against the pilot counts as a regression only when the accepted build passes that same comparison. This audit supplements direct before/after parity, counts and performance checks.

Release qualification also requires --accepted-receipt: a kiln.farm-accepted-build/1 record binding the exact accepted build hash to its owner decision. Without that evidence the runner still produces useful diagnostic results but reports releaseQualified: false. An arbitrary label, reconstruction or adopted count baseline is not appearance acceptance. B-07 records its resource profile explicitly: 39 textures for the prewarmed path and 41 for the current cached-shadow path; legacy pilot allocations remain observations.

For the preserved full-matrix alignment audit, packages/farm/dist/original-draw-after/test aliases the preserved optimized hub binary and alignment-final/test is the historical candidate. Later UI rebuilds retain separate identities and bounded evidence in the alignment plan. The original's 42 files were checked against the historical manifest. Use the shared hardware options above, then run from this directory:

bun scripts/capture-farm-paired.ts --label alignment-farm-paired-local-new-run --accepted-root packages/farm/dist/original-draw-after/test --new-root packages/farm/dist/alignment-final/test --release r34 --backend both
bun scripts/farm-x02-baseline.ts measure --build alignment-final --backends webgpu,webgl2 --out .tmp/alignment-farm/local-new-run-x02-observed.json --log .tmp/alignment-farm/local-new-run-x02-observed-log.json
bun scripts/farm-x02-baseline.ts check --summary .tmp/alignment-farm/local-new-run-x02-observed.json --backend webgpu --json .tmp/alignment-farm/local-new-run-x02-webgpu-check.json
bun scripts/farm-x02-baseline.ts check --summary .tmp/alignment-farm/local-new-run-x02-observed.json --backend webgl2 --json .tmp/alignment-farm/local-new-run-x02-webgl2-check.json

Omit --views on the paired runner to cover all 32 cases; that runner does not accept --views all. The command deliberately supplies no acceptance receipt, so its result remains diagnostic even when every pair passes. Always give count measurement an explicit --out: its default is the committed baseline. A measurement is not authorization to re-baseline. Direct before/after image checks use capture-farm-draw-parity.ts separately; its status now fails for failed, errored or incomplete requested cells and reports that invocation's coverage.

Full-view parity uses the shared luminance and thin-line gates. Historical stream crops retain their explicit luminance-only scope. scripts/timing-ab.ts records GPU clock in MHz and paired busy-times-clock samples with provenance; it is an activity proxy, not energy. Quiet-host/device and paired-decision requirements remain independent of image/count acceptance.