yhz5613813/cadflow/cadflow-step-brep

CadFlow STEP/BREP Inspection and Reverse Engineering

Treat the STEP file as an input artifact to inspect, not as a script to copy. Use the diagnostic namespace outside modeling sessions:

from cadflow.inspect import brep

Read references/inspection-api.md for the exact functions and payload fields.

Choose the task

• Inspect: summarize entities, bounds, volume, area, carrier types, and

topology without constructing a candidate.

• Diagnose: localize differences with sections, entity comparisons,

boundary distances, material regions, or render maps.

• Reconstruct: infer a readable feature sequence and write a new CadFlow

script that does not read the target STEP at runtime.

• Compare: evaluate target and candidate using global properties first, then

strict material or BREP comparisons when the candidate is close enough.

Required workflow

1. Record the target path, file hash, benchmark mode, units/coordinate

assumptions, and output directory.

1. Run inspect_step_rsummary and record validity, body/shell counts, bounds,

volume, area, centroid, and surface/curve statistics.

1. Inspect only the entities needed for the current hypothesis. Use stable

IDs such as face:0 and edge:0; they are deterministic for the same unchanged file but are not semantic IDs across different models.

1. Establish openings, cavities, symmetry, repeated features, section planes,

and likely analytic carriers before choosing primitives or features.

1. Reconstruct with public cadflow APIs. Use import cadflow as cad, keep

the final script independent of the target STEP, and expose inferred dimensions as named parameters.

1. Replay the candidate independently and validate it before comparison.
2. Compare cheaply first: global properties and bounded sections. Use material

difference with include_components=True and no fuzzy tolerance for an equality claim. Use strict BREP comparison only when explicitly required.

1. Save reports, candidate source, replay payload, renders, and comparison

results under the output directory. State what was measured versus inferred.

Inspection boundary

• BREP inspection is diagnostic and must not run inside GraphSession or an

@cad.model callback.

• Do not import private engine modules or direct OCP types in the reconstruction

script.

• Do not search outside the authorized target directory for prior solutions,

caches, deleted files, or another agent's output.

• Do not embed, copy, or re-export target STEP contents into the candidate.

Evidence hierarchy

1. Exact BREP topology identity is the strongest result.
2. Strict bidirectional material equality proves the occupied point set matches,

even if feature histories differ.

1. Matching volume, area, bounds, centroid, and sections are useful gates but

do not prove equality alone.

1. Similar renders are diagnostic evidence only.

Never stop at “looks close” when a better feature order or supported operation could still be tested. If stopping with a structurally different candidate, record the exhausted hypothesis or unsupported operation that justifies it.

Failure handling

• A timeout or fuzzy Boolean leaves equivalence unproved; it is not a pass.
• A large global volume mismatch disproves equality cheaply; do not spend time

on strict difference computation first.

• For a local mismatch, use inspect_section, compare_sections,

compare_boundary_distance, or render_region rather than rerunning every expensive comparison.

• Preserve compact reports and hashes; do not paste huge control-point arrays

into the final narrative.