eli0shin/.dotfiles/forgecad-reconstruct-cad-file

Reconstruct CAD File

The reference asset is evidence, not the deliverable. The deliverable is a readable, parametric .forge.js model that runs, renders, and scores well against the source. Never return Import.mesh()/Import.step() of the source as the final model unless the user explicitly asks for an import wrapper — imports are for measurement, rendering, and scoring only.

Read ../forgecad-build-model/SKILL.md for the sketch-first construction workflow, output policy, and completion criteria. If the user wants to keep the imported file as a component and design around it, use that skill instead of reconstructing the component. For prepared benchmark/RL episodes, follow task-local benchmark instructions. Use ../forgecad/SKILL.md for API/CLI references and ../forgecad/docs/guides/inspection-bundles.md to interpret inspection output.

Workflow

1. Inspect the source directly — the CLI reads CAD/mesh files as inputs, no wrapper script. Gather four evidence types: ls --long stats, an iso render, per-object views, sampled sections (invocations in CLI.md; when sharing a renderer server via --port, run renders sequentially).
2. 3MF: enumerate every build item before scoring — hidden multi-part structure is a common miss. Item ref syntax: Import.mesh docs.
3. Establish construction before modeling: what must be exact vs. approximate vs. parametric; symmetry, origin, controlling sketch planes, extrusion/revolution directions and depths; scoring tolerance and alignment policy. Reason through these choices without generating a brief document.
4. Build the base features first. Use exact sketches and extrude/revolve to match bounds and main geometry, then add functional features and native edge finishes. Recover analytic intent rather than tracing mesh facets.
5. Compare numerically — the core loop:

``bash forgecad compare 3d path/to/source.stl path/to/candidate.forge.js \ --samples 3000 --json --output /tmp/<slug>-score.json ``

1. Iterate coarse to fine: bbox/placement → main volumes/silhouette → holes/bosses/ribs/shells → edge treatments → small details.

Reading the Score

Alignment: start --align none. Use --align center only when origins clearly differ but scale and orientation match. Use --align center-scale only for exploratory diagnosis — it hides dimensional errors.

Calibration: 95+ simple prismatic/revolved parts, 90+ ordinary mechanical parts with fillets/cutouts, 80+ acceptable for organic, faceted, or underdetermined sources. Always report raw rms, p95, max, coverage, bounds delta, and volume delta — a high overall score with a large max can hide a missing local feature.

Faceted sources: decide whether tessellation itself is evidence. Matching low-poly faceting raises the score but reduces parametric clarity — prefer analytic intent unless faceting is part of the artifact or required by the acceptance criteria.

Done Criteria

The final model must run, render, re-compare at --samples 5000, and pass an inspect compare overlay. Open and inspect the overlay and views that expose changed features. Use targeted sections, thickness, fit, or connectivity checks when needed, following ../forgecad-build-model/SKILL.md. Give a short completion message with source and candidate paths, key comparison metrics, and known mismatches classified as intentional simplification or remaining work. Keep temporary scores and inspection output in /tmp under the model-building skill's output policy.