model3d

name	model3d
description	Run an interactive 3D design session — turn a natural-language description (and optional reference images) into print-ready 3D files (STL + 3MF + GLB) in output/, validated for 3D printing and visually verified to match intent. Use whenever the user wants to model, design, generate, mock up, tweak, modify, or 3D-print an object — a bracket, enclosure, stand, holder, knob, gear, fixture, sign, figurine, vase, toy, or any physical part — for Bambu Studio, OrcaSlicer, PrusaSlicer, or MakerWorld. Also use to load and modify an existing STL/3MF/GLB. Triggers on "model a…", "design a…", "make me a printable…", "generate an STL of…", "tweak the…", "I want to 3D print a…".
argument-hint	<natural-language description of the object to model, or a tweak to an existing one>
user-invocable	true
allowed-tools	Read, Write, Edit, Bash, Glob, Agent

/model3d — interactive design session → print-ready 3D

You run a design session: understand the request, build it, look at it, critique it against intent, revise until it both matches the design and is print-ready, then present it. The bundled CLI studio3d (on PATH) does the fabrication, validation, rendering, and history.

Request: $ARGUMENTS (plus any attached images).

Capabilities load on demand (registry). This plugin keeps only model3d and grill-me native to minimize always-on cost. The reference skills and specialist agents this guide names live in the 3d-studio-registry MCP server. When a step says "see the cad-authoring / print-readiness / printer-setup / 3d-modeling-foundations skill", load it with load_skill(id="…"). When it names "the spec-analyst / cad-author / mesh-validator / design-critic subagent", load_agent(id="…") to get its system prompt, then spawn it with the Agent tool. list_skills / list_agents enumerate what's available; studio3d_reference / studio3d_styles give the packaged design grounding.

0. Target the right printer (once per session)

Check the active printer profile so files target the user's real machine:

studio3d profile show        # active profile (bed size, nozzle, AMS, colors)

If there is no active profile, help set one up (see the printer-setup skill): ask make/model, look it up (studio3d printers --search "<model>"), AMS yes/no, and filament colors, then studio3d profile add .... The active profile drives bed-fit, wall minimums, and AMS color mapping automatically.

1. Understand → ModelSpec (+ the design plan, for non-trivial work)

Parse the request (and images) into intent: category (mechanical/functional → CSG; organic/decorative → stylized CSG), key dimensions in mm, and the design cues that define success (for an owl: round body, big eyes, beak, ear tufts; for a phone stand: a cradle that actually holds a phone at a viewing angle). Delegate to the spec-analyst subagent for non-trivial requests or when images are attached. State any dimension you had to assume.

For organic/figurative or otherwise non-trivial requests, base the work on a design plan (design.json) — the source of truth. Run the grill-me session to interview the user and produce a validated plan, or build one directly with studio3d plan new …. Tweaks then edit a plan field and regenerate (step 4).

1b. Look up the packaged REFERENCE + STYLE (organic/figurative)

Before authoring any organic/figurative model, pull the grounded brief and the style params so you build to a known recipe, not by feel:

studio3d reference <subject> --style <style>   # BRIEF: silhouette cues, numeric
                                               # proportions by head-unit H, CSG recipe,
                                               # eye_rule, print_constraints (20 subjects)
studio3d styles <style>                        # head_body_ratio, eye_size_mult,
                                               # feature_exaggeration, facet_level, fillet

Or studio3d plan brief <design.json> to get the brief straight from the plan.

2. Author the geometry (CSG) — BY PROPORTION, with NAMED PARAMETERS

First ground in the knowledge base: studio3d kb "<what you're building>" returns the DFAM numerics + the proven CSG recipe (e.g. "wall bracket gusset", "vessel inner bore wall", "teardrop horizontal hole"). Then write a script in the studio3d DSL (see the cad-authoring skill for the full API: box, cylinder, sphere, ellipsoid, cone, prism, tube, slot, teardrop, extrude, revolve, twist_extrude, loft, hull, text, interference, arrange_on_bed + booleans + transforms).

Parameterize every key dimension by reading it from the injected P dict with a default — wall = P.get("wall", 2.0) — NOT magic numbers. The bundle ships model.py

• params.json, so the model is forever-editable (studio3d tweak --set wall=3 regenerates it deterministically). A hardcoded dimension is a defect.

For organic/figure models, author by proportion: define a head-unit H from the reference recipe and place/size every part as a multiple of H (per the brief's numeric ratios and eye_rule) — not ad-hoc mm. Apply the print-readiness rules while modeling (mm, ≥0.8mm walls, ≥45° overhangs, base chamfers, teardrop horizontal holes, mating clearance). Delegate complex parts to the cad-author subagent. End with .on_bed().

For multi-part assemblies, build each part, lay them out with arrange_on_bed([...]), and assert no collision with interference(a, b) == 0 before returning — then capture the mates + one clearance_mm knob in the design plan's assembly block.

3. THE FIDELITY LOOP — fabricate, look, critique, revise (≤4 passes)

This is how the result actually matches the request. Do not ship a model you have not looked at. It is vision-grounded: you compare renders against BOTH the subject's reference silhouette cues AND any 2D images the user gave you.

# from a plan (preferred for organic/figure work):
studio3d gen-script --script model.py --plan <slug>.design.json --out output
# or ad-hoc:
studio3d gen-script --script model.py --name "<slug>" --prompt "$ARGUMENTS" \
  --category <cat> --color "<hex>" --out output
studio3d render output/<slug>/model.stl --color "<hex>" --views front,right,top,iso

Pull the ground truth, then Read the rendered views (output/<slug>/views/view_*.png) and the user's reference images side by side:

studio3d reference <subject> --style <style>   # silhouette cues + numeric proportions

Score each criterion 0–100 against that ground truth:

Criterion	Scored against
Silhouette	Do the renders read as the subject's silhouette cues / the user's images?
Proportion	Do part ratios match the reference head-unit (H) ratios?
Feature	Is every reference cue present + identifiable (eyes, beak, ear tufts…)?
Style	Does it match the style params (head_body_ratio, eye_size_mult, facet_level)?

Also read the kernel metrics in report.json → metrics.kernel_metrics (or the summary.kernel_metrics from gen-script): n_components must be 1 for a single part (2+ = a floating/disconnected piece a render can hide), genus as expected, and wall_p05_mm ≥ the minimum. Vision alone is contextually blind — fuse it with these numbers. Require print-readiness (print_ready: true, no blocking D1/D3).

Revise the script for any low score and regenerate (same --plan/--name → overwrites in place, git captures the change). Stop when all axes are ≥ ~95 AND print_ready is true, or after 4 passes keeping the best-so-far. If the SAME issue persists across two passes, do NOT repeat the fix — escalate: switch to a fundamentally different construction (decompose into sub-parts, hull/loft for blobby forms, change the proportion anchor).

For a fresh, frame-isolated, STRONGER judgement, delegate scoring to the design-critic subagent (it runs on a strong model, renders + critiques blind, fuses the kernel metrics, and returns a rubric score + fixes, with a ship/revise/escalate/reject verdict).

4. Tweak / modify / iterate

The user will refine ("make it taller", "add a lid", "round the corners"). When the model has a design plan, edit the relevant field in design.json (it is the base design), bump revision, re-plan validate, and regenerate from the same plan. Either way edit the same script and regenerate with the same --name/--plan so it overwrites in place (git history preserves every revision — studio3d history --bundle <slug>; recover with studio3d history --bundle <slug> --revert <sha>). Only pass --variant, or branch to a new design.json, when the user wants a separate copy.

5. Modify an EXISTING model file

If the user provides an STL/3MF/GLB:

studio3d import path/to/their.stl --name <slug> --reorient --out output

Then modify it by authoring a script that calls load_mesh("output/<slug>/model.stl") and applies booleans/transforms (cut a hole, add a mount, scale, emboss), and run it through gen-script — the fidelity loop applies the same way.

6. Present + live preview

Report what was built, the intent-match and print-readiness scores, the bundle path, the recommended slicer format (3MF for Bambu/AMS color, STL otherwise), key metrics (mm, mass), and print advice (supports/orientation). Mention that the bundle ships an auditable Print-Readiness Certificate (certificate.json: prompt → script hash → D1–D4 → slice → human approval) and editable source (model.py + params.json).

For multicolor / AMS, author parts with .paint("#hex") and combine via multicolor_union(...) — the bundle's model.3mf then carries per-face AMS colors and the GLB shows them; the manifest marks multicolor + palette. For CAD interchange, add step to --formats (or run studio3d step <bundle>/model.stl) to emit a faceted STEP solid that opens in Fusion/FreeCAD/SolidWorks.

Real extras the user can run:

• studio3d slice <bundle>/model.3mf — REAL slice-to-G-code (D2 ground truth): print time + filament grams (needs a slicer installed; otherwise D2 is the labeled proxy).
• studio3d orient <bundle>/model.stl --out reoriented.stl — support-minimizing pose.
• studio3d step <bundle>/model.stl — faceted STEP for CAD interchange.
• studio3d tweak --plan <slug>.design.json --set <k>=<v> --script model.py — change one knob and regenerate deterministically.
• studio3d certify <bundle> --approve — record human sign-off in the certificate.

The user can watch every change live in the viewer (3D view, print-readiness report, and a parametric Customizer that surfaces the knobs):

cd web && npm install && npm run dev    # http://localhost:5173 — auto-refreshes

Guardrails

• Always millimeters. Manifold is non-negotiable (the CSG engine guarantees it; if a result isn't watertight, a boolean was degenerate — nudge a dimension by 0.01mm).
• Always look before you ship (step 3). Validation proves it prints; the render proves it matches. Both must pass.
• Prefer the local CSG path; organic shapes are authored as stylized CSG by default.
• Don't silently invent safety- or fit-critical dimensions — confirm.