Files
make-cad/docs/skill/references/workflows/step-reconstruction.md
T
NiJingzhe cdfc0ee904 docs(skill): re-studio workflow rewritten for the composer round
- reverse-engineering-studio.md: free-form composer (entity/op chips,
  [label](token) serialization, IME-safe Enter), operation tips as a
  markdown registry (viewer/server/operations/, mtime-hot, errors named),
  submission schema incl. operations + operation_context, self-contained
  artifact rule (transcription = literals, extraction harnesses beside
  the case), kernel-ordering facts (bores before blend fillets; fitter
  to_dict unique_knots contract), artifact names the UI polls
- step-reconstruction.md: route to the studio variant when a human
  operator is available
- SKILL.md: workflow count corrected to six; viewer/README.md updated
  for the composer + tips registry
2026-09-05 14:35:20 +08:00

5.5 KiB

Workflow: STEP Reconstruction

Rebuild an editable SimpleCAD model from an existing STEP file, with evidence gathering, feature-tree inference, candidate construction, and tiered acceptance.

Goal and scope

Use when the deliverable is an editable/replayable model reproducing a target .step/.stp. Not for read-only questions about a STEP file — those stay in domains/step-inspection.md without reconstruction. When a human operator is available to annotate feature intent on the target, run the variant workflow references/workflows/reverse-engineering-studio.md instead of this autonomous path.

Task decomposition

goal: candidate model accepted against the target STEP
primary_domain: step-inspection
required_domains:
  - requirement-refinement
  - step-inspection
  - part-modeling
optional_domains:
  - sketch-and-features
  - assembly-and-product     # multi-component targets
  - export-and-translation   # re-export for comparison or delivery
artifacts:
  - target evidence notes (summaries, entity descriptors, sections)
  - feature-tree hypothesis with provenance rows
  - candidate source and replayed model
  - comparison evidence and acceptance verdict
validation_gates:
  - replay succeeds in a fresh process
  - STEP roundtrip of the candidate validates
  - acceptance follows the hierarchy (topology identity > float drift > justified stop)
repair_routes:
  - comparison fails -> step-inspection (localize) then part-modeling (fix feature)
  - global similarity but locally falsified feature -> change the feature hypothesis

Steps

  1. Refine the requirement: what fidelity is required (exact transcription vs functional equivalence), which components are in scope, what acceptance means here.

  2. Inventory the target economically (domains/step-inspection.md): inspect_step_rsummary (with include_parameter_groups=True once), entity descriptors for specific hypotheses, compact boundaries/sections. Persist large arrays to files, not conversation context. Read references/inspect/brep-reverse-engineering.md completely first.

  3. Infer feature families from carrier and section behavior (references/docs/guides/reconstruction-agent-strategy.md):

    shared axis + rotational sections       -> revolve
    stable translated profile               -> extrude
    profile transported along a path        -> sweep
    ordered changing section family         -> loft
    verified equal angular/linear units     -> one unit + pattern
    several local signatures                -> ordered mixed feature tree
    

    Record a provenance row per region: observed boundary, adjacent-carrier evidence, proposed operation, direction/axis, confidence, and the check that could falsify it. A loop on a final planar face may be a later hole or pocket, not part of the base profile.

  4. Build the candidate via single-part-modeling.md (block structure per discipline/feature-tree-convention.md; sketch tier when constraints represent the evidence (domains/sketch-and-features.md), direct geometry tier for non-planar or exact transcription, annotated profile=geometry honestly). Overshoot through-cut tools; validate one representative feature before patterning.

  5. Replay in a fresh process: export_model_json → replay_model_json; then validate_step_roundtrip_rdescriptor on the candidate's STEP export.

  6. Compare and iterate (domains/step-inspection.md): compare_global_properties_rdescriptor → scoped compare_boundary_distance_rdescriptor (≤200 samples, face-id scoped) → compare_material_rdescriptor(include_components=True) for strict proof → compare_steps_rbrepcomparison for the exact gate. Use track_section_contours_rdescriptor when section topology (birth/death/split/merge) is the open question, fit_face_analytic_rdescriptor to test carrier families, render_step_comparison_rpath for shared-view visual evidence. After a failed strict comparison, group causes with BRepComparison.to_error_summary(), fix related groups together, rerun a fresh cycle.

  7. Accept per the hierarchy (domains/step-inspection.md): topology identity is the endpoint; float-level serialization drift is acceptable; a structurally different but visually close stop requires recorded evidence that no better feature order exists or the SDK lacks the operation.

Validation gates

  • One compact iteration row per attempt: hypothesis, changed parameters, replay/roundtrip status, diagnostic deltas, strongest proof attempted.
  • Preserve the best valid candidate; repeated non-improvement changes the feature hypothesis instead of blind parameter tuning.
  • Global similarity never overrides a locally falsified feature.

Failure routes

  • No affordable evidence for a region → record the gap; do not invent geometry and call it recovered.
  • Candidate structurally different → re-open the feature-tree hypothesis before accepting a visual-close stop.
  • Reconstruction tooling absent (inspect extra not installed) → report the missing capability rather than degrading silently.

Deliverables

Candidate source path, replay evidence, comparison evidence with verdict and its acceptance tier, provenance rows, and explicitly unrecovered regions.

Required reading before authoring or editing any part source in this workflow: discipline/feature-tree-convention.md — the block-structured sketch → basic body op → bool → modifier convention, its mandatory boundary comments, and the sketch/geometry/primitive tier rules.