Control the Serpentine3D NURBS 3D modeller. Objects are referenced by name (e.g. 'Curve 01') or id. Typical flow: serp_scene_info to see what exists, create geometry, then serp_screenshot to look at the result. Units are generic model units; the construction plane is world XY with Z up.
15 tools with mostly complete schemas and descriptions, but significant gaps in parameter documentation and error handling guidance. Tool names follow verb_noun convention well (create_curve, serp_transform, serp_select). Descriptions average ~150 chars, within baseline range (p10=34, p90=392). However, many parameters lack descriptions or have vague ones. serp_command is a catch-all that violates single-responsibility principle. No output schemas documented. Error recovery guidance absent. Composition is reasonable, tools chain well via object names, but parameter relationships underdocumented.
Boolean operation between solids: 'union', 'difference' (targets minus tools), or 'intersection'. Tools are consumed.
Run any Serpentine3D command exactly as if typed in the command line, supplying its interactive inputs in order. Examples: command='circle', inputs=['0,0,0', '5']; command='extrude', inputs=['Curve 01', '', '20', 'Yes'] (object name(s), then '' to end selection, then distance, then cap); command='zoomextents'. Selection prompts accept object names, 'all', or '' to finish.
Create a NURBS curve from 3D points [[x,y,z], ...]. kind: 'interp' (curve passes through the points), 'control' (points are control vertices), 'polyline' (straight segments), 'line' (two points).
Create a surface/solid from existing curves (referenced by name). operation: 'extrude' params: direction [x,y,z] (default [0,0,1]), distance (default 10), cap (bool, default true - closed profiles become solids); 'revolve' params: axis_point, axis_dir, angle (degrees, default 360); 'loft' curves: 2+ profiles in order; params: ruled (bool); 'planar' flat surface from one closed planar curve; 'sweep' curves: [profile, rail]
Export the scene (or current selection) to .serp, .step/.stp or .obj — format chosen by extension.
serp_command violates single-responsibility principle. It accepts any command + inputs, making it a catch-all that bypasses structured tools. LLMs cannot infer when to use it vs. specific tools, leading to poor tool selection and reduced composability.
serp_create_surface and serp_transform have vague 'params' objects (type: object, no properties defined). LLMs cannot validate inputs or understand required/optional fields. Operation-specific parameters (direction, axis_point, angle) are documented in description text, not schema.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., what does serp_scene_info return? object names? layer structure?). Breaks tool chaining and forces agents to guess response structure.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2026-07-28+ | v2 |
Import a file into the scene (.serp, .step/.stp, .obj).
Manage layers. action: 'list', 'create', 'rename', 'visible', 'current', 'color' (color as [r,g,b] 0-1), 'assign' (move objects to layer), 'delete'. A layer can live under another: pass parent to 'create', and name a layer by its path ('Walls::Interior') where two layers share a name.
Measure geometry. what: 'distance' (needs points=[[..],[..]]), 'length', 'area', 'volume', 'bbox', 'centroid' (need targets).
Open Python in a new editable Script draft without executing it. Existing drafts and geometry are preserved. Scripts can use doc (a working document), geo (geometry helpers), and selected (selected SceneObjects). The user can edit the draft, Run to preview, then Keep or Discard its changes.
Get the current scene: objects (name/kind/layer/bbox), layers, selection, bounds, and display mode.
Capture the current 3D viewport so you can see the model. Returns a PNG image. Use after making changes to check the result.
Select objects by names, kind (curve/surface/solid), or layer name. mode: 'replace', 'add', or 'clear' (clear ignores other args).
Transform objects (by name). operation/params: 'move' offset [dx,dy,dz]; 'copy' offset [dx,dy,dz]; 'rotate' center [x,y,z], axis [x,y,z] (default Z), angle degrees; 'scale' center, factor (uniform) or factors [sx,sy,sz]; 'mirror' plane_point, plane_normal, keep_original (bool)
Undo the last operation (or redo with redo=true).
Adjust the viewport. view: top/front/right/left/back/bottom/isometric/perspective. display_mode: wireframe/shaded/ghosted. zoom_extents fits all objects in view.
serp_layers action parameter has 8 enum values but many are underdocumented. 'assign' requires 'objects' param; 'create' requires 'parent' param. Dependencies between parameters not stated in descriptions, risking misuse.
No error handling guidance. Tools return success/failure but do not tell LLMs what to do on failure (retry? ask user? abort?). No recovery paths documented (e.g., 'If import fails, check file path format').