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.

3.2 KB
Raw Blame

Muse optical instrument: creative refinement

A separate fresh OpenCode conversation refined the supplied Muse Q1 source using opencode/muse-spark-1.3-contributor-free. It used the local 1d3bf676 package, project-local author/refine/QA skills and an explicit Node 22.23.1 executable and PATH. This is provided-source curation, not independent authoring or a pilot rerun. The brief allowed a substantial redesign and specified a 40-call / 64-image-cell / 15-minute budget. No paid API fallback or additional attempt was used.

The harness exited 0 after 520 seconds. The observer recorded seventeen calls and seventeen image cells. The model reported fifteen calls and twenty-four cells; the observer counters take precedence. The final source SHA-256 is 32223e1d94c3034d14f5d3f401fd7fbb5da3e471ddf92c3acdc319a85a03106d, matching its reference. The exported GLB SHA-256 is 1c5def38ea152a30d5afee13c958eeed05858c7057e9048a0379fd515a820aa4.

The model replaced the white tube and tripod with a dark teal lofted pedestal, segmented barrel, swept ribs, recessed optical opening, focusing slide and small adjustment hardware. It then edited the bore floor and baffles by reference. Actual returned images include whole-asset, attachment and axial opening views, plus two locked-camera motion strips. Source and GLB were exported from the same final reference through the installed CLI.

The first redesign failed with the generic execution-rejected message. Validation still returned true. Small independent probes isolated a rejected gearGeo call; the model replaced it with a disc and individual tooth boxes. Post-run source inspection explains that probe: it set tipRadius: 0.075 while omitting rootRadius, whose default is 0.8. That invalid radius ordering should produce a specific repair hint. The generic error instead encouraged the model to conclude that gears were unsupported. A separate gear geometry defect was found during follow-up engine investigation; it does not change this probe’s invalid inputs. The earlier failed revision and probes remain in the run record.

The final six-view sheet is a coherent gallery study, with more deliberate massing than the initial Q1 telescope. It is not an exceptional hero asset. The brass highlights are strong, the large pedestal remains plain, and the recessed lens shows smeared studio reflections. The model's claim that every final view is good is not adopted as an independent judgment. Motion samples show the focusing slide moving while attached; hidden clearances and real optical performance are untested.

The exported source is retained byte-for-byte as examples/bench-refractor.kiln.js. Its public sidecar separates original authoring from this supplied-source curation. A site rebuild and fresh poster must be reviewed before selecting it for the home page. The source is not hand-polished to inflate the model result.

Subsequent renderer investigation found that the original GPU image path clipped bright surfaces before applying output color conversion. The old fidelity receipt is retained as observed, but those images are not definitive material-quality evidence. Final gallery selection uses corrected renders of the downloadable GLB.