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.

3.4 KB
Raw Blame

Muse / OpenCode: installed-package journey

On 5 September 2026, opencode/muse-spark-1.3-contributor-free through OpenCode 1.18.27 completed the optical-instrument journey in a fresh conversation and workspace. The run used a local npm installation of candidate 614b1c9ea2adf55bd7b4c1b058c0b7c7f4dd24f76f8a2d22267c6fb2221281ef, installed without checkout source or development dependencies, and project-local author, refine and QA skills. No completed example was supplied. External-directory access was denied. This is one workflow observation, not a model ranking.

The harness exited successfully after 216 seconds. The transparent observer recorded eleven Kiln requests and twelve image cells, within the declared 30-call / 48-cell / 15-minute budget. Four grouped discovery calls returned 24 named signatures. After the first 10,499-byte program, later tool calls used its reference; none repeated the full program.

Independent file checks confirm these exported source hashes:

  • First: 12c5e05e4e2fe8996831fa0850d885fb8b4b8e92e6929e5000be2476c16067d8.
  • Final: f21fc9f7ab93dfe56720232007668ba95fc541d6071454a7ba12289c5a519df0.
  • Final GLB: b02204523825ef306f27a1b3b660d62459b538d97555d88ebbf4c0cc01a287d8.

The source difference is exactly one line: serviceOffset changed from 0.05 to 0.20. Every other source byte is unchanged. The model exported the final source, GLB and image sheet through the installed CLI. The run did not export a first GLB, so there is no independent before/after GLB byte comparison.

The observer saved the actual initial six-view grid, a part inspection, the revised whole/detail grid and a three-frame animation strip. Independent visual review of the revised images confirms a coherent telescope and a carriage that moves and returns in the locked view. It does not establish hidden clearance. The model explicitly reported the 1.76 m height exceeds the 1.5 m brief and that concealed bolts were not individually imaged. Its material-fidelity response said full-material, non-degraded, with exactArtifact: false; the image is not claimed to depict the exact exported GLB bytes.

The design is a plain narrow tube on a tripod with a dark capped-looking opening. It passes a useful source-reuse journey but does not meet the lead-gallery visual bar. A subsequent creative refinement is a separate provided-source task with its own budget and provenance; it cannot replace this initial result in the record.

Runtime qualification

The absolute MCP launcher resolved to Node 24.14.1, despite the interactive shell announcing Node 22.23.1. The project CLI used node from its inherited shell; that child process version was not separately logged. Installation used npm 11.11.0 and reported the expected engine-version warning. Therefore this run demonstrates a working local installation on that actual route, not a pinned Node 22.23.1 reproduction. Future curation explicitly pins the Node executable and PATH.

The later 1d3bf676 candidate changes cancellation outcome ordering and docs; this successful 614b1c9e journey does not recertify that negative branch. Offline cancellation tests and final-package installation checks cover it separately.

Later GPU investigation identified missing output conversion that clipped bright surfaces in these original images. The reported fidelity fields are historical receipts; material judgments require corrected rendering. Source edits, geometry and animation observations remain separately recorded.